Beyond Thermodynamics: Tackling the Fundamental Challenges in Predicting Synthesizable Inorganic Crystals

Abstract

The acceleration of computational materials design has starkly contrasted with the slow, empirical nature of experimental synthesis, creating a critical bottleneck in materials discovery. This article explores the fundamental challenges in predicting the synthesizability of inorganic crystals, moving beyond traditional proxies like thermodynamic stability. We examine the limitations of conventional methods and the rise of advanced machine learning solutions, including deep learning models and large language models, which learn synthesizability rules directly from experimental data. The scope covers foundational concepts, methodological innovations for practical application, strategies to overcome data and model limitations, and rigorous validation of these new approaches. For researchers and development professionals, this synthesis provides a crucial guide to navigating the transition from virtual candidate to synthetically accessible material, a transformation with profound implications for the development of new functional materials.

Why Synthesizability is a Grand Challenge in Materials Informatics

Advancements in computational chemistry and materials science, particularly in crystal structure prediction (CSP), have revolutionized the virtual design of new materials with targeted properties [1]. High-throughput computational screening and AI-powered generative models can now propose millions of hypothetical candidate materials. However, a critical bottleneck persists: the vast majority of these computationally designed materials, despite being thermodynamically stable, are not synthesizable [1] [2]. This creates a fundamental gap between theoretical predictions and experimental realization, severely limiting the impact and throughput of materials discovery pipelines.

The core challenge lies in the complex nature of synthesizability itself. Unlike thermodynamic stability, which can be reasonably estimated from first principles, synthesizability encompasses kinetic factors, precursor selection, reaction pathways, and specific experimental conditions—most of which cannot be fully predicted based on thermodynamic or kinetic constraints alone [1] [3]. This complexity is compounded by the fact that experimental synthesis reports (positive data) are documented in scientific literature, while failed synthesis attempts (negative data) are rarely reported, creating a fundamental data limitation for machine learning models [3] [2]. This article examines the fundamental challenges in predicting synthesizable inorganic crystals and explores the integrated computational and experimental strategies being developed to overcome the synthesis bottleneck.

Current Computational Paradigms for Predicting Synthesizability

Thermodynamic Stability and Its Limitations

Traditional computational materials discovery has relied heavily on thermodynamic energy-based stability predictions as a proxy for synthesizability. Density Functional Theory (DFT) calculations are used to determine a material's formation energy and assess whether it is stable against decomposition into competing phases [3]. While necessary for identifying stable materials, this approach has proven insufficient for predicting synthesizability. A significant limitation is that many thermodynamically stable materials remain unsynthesized, while many metastable materials (those not at the global energy minimum) can be successfully synthesized through kinetically controlled pathways [1] [2].

The performance of formation energy calculations as a synthesizability filter is quantitatively limited; they capture only approximately 50% of synthesized inorganic crystalline materials [3]. Similarly, the commonly employed charge-balancing criterion—filtering materials based on net neutral ionic charge—also shows poor performance, with only 37% of known synthesized materials satisfying this constraint [3]. These findings underscore that synthesizability depends on factors beyond simple thermodynamics or charge neutrality.

Data-Driven Machine Learning Approaches

Machine learning (ML) models trained on databases of known materials have emerged as powerful tools for synthesizability prediction. These approaches can be broadly categorized by their input data type and methodological framework.

Table 1: Comparison of Machine Learning Approaches for Synthesizability Prediction

Model Type	Input Data	Key Advantages	Limitations	Representative Models
Composition-Based	Chemical formula only	Applicable when structure is unknown; fast screening	Cannot differentiate between polymorphs	SynthNN [3]
Structure-Based	Crystal structure	Accounts for structural polymorphs; higher accuracy	Requires predicted structure	PU-CGCNN [2], Synthesizability-driven CSP [1]
Positive-Unlabeled (PU) Learning	Composition or structure	Handles lack of negative data; realistic for materials space	Complex training and evaluation	Various implementations [3] [2]
LLM-Based	Textual structure descriptions	Human-interpretable; explainable predictions	Dependent on description quality	StructGPT, PU-GPT-embedding [2]

Composition-based models like SynthNN (Synthesizability Neural Network) operate solely on chemical formulas, making them applicable for screening materials where atomic structure is unknown. These models learn chemical principles directly from data distributions, implicitly capturing relationships like charge-balancing, chemical family tendencies, and ionicity without explicit programming [3]. Remarkably, in benchmark tests, SynthNN achieved 1.5× higher precision in identifying synthesizable materials compared to the best human experts and completed the task five orders of magnitude faster [3].

Structure-based models leverage the atomic arrangement of crystal structures, enabling them to differentiate between polymorphs of the same composition—a critical capability given the prevalence of polymorphic materials like diamond and graphite [1]. These models utilize various structure representations, including graph-based encodings, Fourier-transformed crystal features, and Wyckoff position encodings [1] [2].

The Positive-Unlabeled (PU) learning framework addresses a fundamental data challenge in synthesizability prediction. Since only synthesized materials (positive examples) are definitively known, while unsynthesized materials constitute an unlabeled set (which may contain both synthesizable and non-synthesizable materials), PU learning provides appropriate mathematical foundations for model training [3] [2].

Recent advances incorporate Large Language Models (LLMs) fine-tuned on text descriptions of crystal structures generated by tools like Robocrystallographer. These models not only achieve performance comparable to traditional graph-neural networks but also provide human-readable explanations for their predictions, enhancing interpretability [2]. The LLM-based workflow can generate explanations for the factors governing synthesizability and extract underlying physical rules to guide chemists in modifying non-synthesizable hypothetical structures [2].

Performance Benchmarking of Computational Approaches

Table 2: Performance Comparison of Synthesizability Prediction Methods

Method	True Positive Rate (Recall)	Approximated Precision	Key Strengths	Test Conditions
DFT Formation Energy [3]	~50%	Not specified	Strong theoretical foundation	Captured % of known synthesized materials
Charge-Balancing [3]	37% (all inorganics), 23% (Cs compounds)	Not specified	Chemically intuitive, fast	Percentage of known materials that are charge-balanced
SynthNN (Composition-Based) [3]	Not specified	7× higher than DFT	High throughput; learns chemical principles	Comparison against human experts
PU-CGCNN (Structure-Based) [2]	Baseline	Baseline	Established structure-based benchmark	MP30 dataset with α-estimation
StructGPT (LLM-Based) [2]	Comparable to PU-CGCNN	Comparable to PU-CGCNN	Explainable predictions	Same as above with GPT-4o-mini
PU-GPT-Embedding [2]	Highest among compared methods	Highest among compared methods	Combines LLM representations with PU-classifier	Same as above

The performance advantages of machine learning approaches are evident in these comparisons. The integration of structural information typically improves prediction quality over composition-only models, while LLM-based approaches offer additional benefits in interpretability and potential cost efficiency [2].

Experimental Validation and Synthesis Planning

Autonomous Synthesis Platforms

The translation of computational predictions to synthesized materials is increasingly being automated through autonomous synthesis platforms. These systems integrate robotics, real-time analytics, and synthesis planning algorithms to execute multi-step synthesis of inorganic materials with minimal human intervention.

The hardware infrastructure for such platforms typically includes:

• Liquid handling robots and robotic grippers for precise transfer of reagents and vessels
• Computer-controlled heater/shaker blocks for reaction execution
• Automated purification systems for product isolation
• Analytical instrumentation (e.g., LC/MS, NMR) for product verification and quantification [4]

A significant engineering challenge involves developing universally applicable purification strategies that can be automated. While specialized approaches like Burke's iterative MIDA-boronate coupling platform use catch-and-release methods for specific reactions, a general purification strategy for diverse inorganic materials remains elusive [4].

Synthesis Route Prediction

Predicting viable synthesis routes represents a complementary approach to synthesizability assessment. Recent work has developed element-wise graph neural networks to predict inorganic synthesis recipes from target compositions [5]. These models outperform popularity-based statistical baselines and demonstrate temporal validity—when trained on data until 2016, they successfully predict synthetic precursors for materials synthesized after 2016 [5].

The probability scores generated by these models correlate with prediction accuracy, serving as useful confidence metrics that enable experimentalists to prioritize synthesis attempts [5]. This capability is particularly valuable for resource-intensive solid-state synthesis experiments.

Integrated Workflows and Case Studies

Synthesizability-Driven Crystal Structure Prediction

A promising framework for bridging the synthesis gap integrates synthesizability evaluation directly into the crystal structure prediction process. This synthesizability-driven CSP approach combines symmetry-guided structure generation with machine learning-based synthesizability assessment [1].

The methodology involves three key stages, illustrated in the workflow below:

This workflow successfully reproduced 13 experimentally known XSe (X = Sc, Ti, Mn, Fe, Ni, Cu, Zn) structures and identified 92,310 potentially synthesizable structures from the 554,054 candidates initially predicted by the GNoME (Graph Networks for Materials Exploration) project [1]. The approach also predicted three novel HfV₂O₇ phases with low formation energies and high synthesizability, demonstrating its potential for discovering viable synthesis targets [1].

Explainable AI for Synthesis Guidance

A significant advancement in synthesizability prediction is the development of explainable AI approaches that not only predict synthesizability but also provide human-understandable reasons for the predictions. As illustrated below, LLM-based workflows can generate structural descriptions and synthesizability explanations:

This explainable AI framework helps materials scientists understand the structural or chemical features that contribute to synthesizability challenges, enabling more informed design of synthesizable materials [2].

• Materials Project (MP) Database: Contains computational and experimental data for over 150,000 materials, including crystal structures, formation energies, and band structures. Essential for training ML models and benchmarking predictions [1] [2].
• Inorganic Crystal Structure Database (ICSD): Comprehensive collection of experimentally determined inorganic crystal structures. Serves as the primary source of "positive" examples for synthesizability models [3].
• Robocrystallographer: Open-source toolkit for generating text-based descriptions of crystal structures. Converts CIF files into natural language descriptions for LLM-based prediction approaches [2].
• Open Reaction Database: Emerging source of standardized chemical reaction data, including synthesis procedures and outcomes. Addresses critical data limitations in synthesis planning [4].

Experimental Infrastructure

• Automated Synthesis Platforms: Integrated systems like the Chemputer or Eli Lilly's automated synthesis platform that translate digital synthesis recipes into physical operations. Enable high-throughput experimental validation [4].
• Advanced Characterization Suite: Combination of analytical techniques including LC/MS for separation and identification, NMR for structural elucidation, and corona aerosol detection (CAD) for universal quantitation without standards [4].
• Chemical Inventory Management: Automated storage and retrieval systems for maintaining extensive collections of precursors and building blocks. Critical for enabling diverse synthesis campaigns [4].

The synthesis bottleneck represents a fundamental challenge in materials discovery that intersects computational prediction, experimental synthesis, and data science. While significant progress has been made in developing computational models that surpass human expert performance in identifying synthesizable materials, fully bridging the gap between virtual design and real-world materials requires continued advancement in several key areas.

The integration of explainable AI approaches will be crucial for building trust in predictive models and providing actionable insights for materials design. Furthermore, the development of universal purification strategies and more sophisticated synthesis route prediction algorithms will enhance the feasibility of autonomous materials synthesis. As these technologies mature, the vision of fully autonomous materials discovery pipelines—from computational design to synthesized and characterized materials—comes closer to reality, promising to accelerate the development of next-generation materials for energy, electronics, and beyond.

The convergence of synthesizability-driven CSP, explainable AI, and autonomous synthesis platforms represents a paradigm shift in materials discovery, one that ultimately transforms the synthesis bottleneck from a formidable barrier into a manageable engineering challenge.

The discovery of new inorganic crystalline materials has been revolutionized by computational methods, particularly density functional theory (DFT), which can screen millions of hypothetical compounds for desirable properties. However, a critical bottleneck persists: the significant disparity between computationally predicted materials and those that can be successfully synthesized in the laboratory. For decades, thermodynamic stability, typically assessed through formation energy and energy above the convex hull, has served as the primary proxy for synthesizability. This paradigm operates on the assumption that compounds with favorable formation energies are synthetically accessible, while those with unfavorable energies are not. Yet, this assumption fails to account for the complex kinetic and experimental factors that govern actual synthesis outcomes. This whitepaper examines the fundamental limitations of relying solely on thermodynamic stability for synthesizability prediction and explores the emerging data-driven approaches that are bridging this critical gap.

The inadequacy of traditional metrics is quantitatively evident. While thermodynamic stability methods can identify compounds unlikely to decompose, they incorrectly label many metastable-yet-synthesizable materials as non-synthesizable while also missing numerous energetically favorable compounds that have never been synthesized. This discrepancy arises because synthesis is governed not only by thermodynamic driving forces but also by kinetic pathways, precursor selection, reaction conditions, and experimental feasibility constraints that transcend simple thermodynamic considerations [6]. The development of accurate synthesizability predictors therefore represents a fundamental challenge in the field of computational materials design, one that must account for the complex interplay of multiple physical and chemical factors beyond bulk thermodynamic stability.

The Quantitative Shortcomings of Thermodynamic Stability Metrics

Traditional approaches for identifying promising synthesizable materials typically involve assessing thermodynamic formation energies or energy above convex hull via DFT calculations. The underlying premise is that materials with negative formation energies and small or positive energies above the convex hull are thermodynamically stable or metastable and thus potentially synthesizable. However, numerous structures with favorable formation energies have yet to be synthesized, while various metastable structures with less favorable formation energies are successfully synthesized [7]. This fundamental disconnect reveals the limitations of thermodynamic stability as a comprehensive synthesizability metric.

Table 1: Performance Comparison of Synthesizability Prediction Methods

Prediction Method	Key Metric	Accuracy/Performance	Principal Limitation
Thermodynamic Stability (Formation Energy)	Energy above hull ≥0.1 eV/atom	74.1% accuracy [7]	Misses metastable phases; ignores kinetic factors
Kinetic Stability (Phonon Spectrum)	Lowest frequency ≥ -0.1 THz	82.2% accuracy [7]	Computationally expensive; imaginary frequencies don't preclude synthesis
Charge Balancing	Net ionic charge neutrality	37% of known compounds charge-balanced [3]	Overly simplistic; fails for metallic/covalent systems
SynthNN (Composition-based ML)	Precision in discovery	7× higher precision than formation energy [3]	Lacks structural information
CSLLM Framework (Structure-based LLM)	Overall accuracy	98.6% accuracy [7]	Requires structural input; training data limitations

The performance gap between thermodynamic metrics and modern machine learning approaches is striking. Large Language Models (LLMs) fine-tuned for synthesizability prediction, such as the Crystal Synthesis LLM (CSLLM) framework, achieve 98.6% accuracy in testing, significantly outperforming traditional thermodynamic screening based on energy above hull (74.1%) and kinetic stability assessment via phonon spectrum analysis (82.2%) [7]. Similarly, the SynthNN model demonstrates 7× higher precision in identifying synthesizable materials compared to DFT-calculated formation energies [3]. These quantitative comparisons underscore the severe limitations of relying solely on thermodynamic stability for synthesizability assessment.

Fundamental Limitations of the Thermodynamic Paradigm

The Metastability Challenge

Thermodynamic stability metrics fundamentally fail to account for metastable materials that can be synthesized through kinetic control. Many experimentally realized compounds are metastable under standard conditions but become accessible through specific synthesis pathways that bypass thermodynamic equilibrium. For instance, in the La-Si-P ternary system, computational insights reveal that thermodynamic stability alone cannot explain the synthetic challenges encountered for predicted ternary phases (La₂SiP, La₅SiP₃, and La₂SiP₃). Molecular dynamics simulations using machine learning interatomic potentials indicate that the rapid formation of a Si-substituted LaP crystalline phase creates a kinetic barrier to synthesizing the predicted ternary compounds, despite their computed thermodynamic stability [6]. This exemplifies how kinetic competition between phases, rather than thermodynamic stability alone, governs synthetic accessibility.

Beyond Bulk Energetics: The Role of Synthesis Environment

Traditional thermodynamic approaches typically consider only the bulk energetics of the starting materials and final products, ignoring the complex reaction pathways and environmental factors that dictate experimental synthesis. The synthesis process is influenced by numerous factors beyond bulk thermodynamics, including precursor selection, reaction kinetics, temperature profiles, pressure conditions, and the presence of catalysts or flux agents. These factors collectively determine whether a thermodynamically favorable compound can actually be synthesized [6]. Phase diagrams offer a more direct correlation with synthesizability as they delineate stable phases under varying temperatures, pressures, and compositions. However, constructing the free energy surface for all possible phases as a function of these variables is computationally impractical for high-throughput screening [7].

Emerging Approaches Beyond Thermodynamic Stability

Machine Learning with Positive-Unlabeled Learning

Machine learning approaches, particularly those employing Positive-Unlabeled (PU) learning frameworks, have emerged as powerful alternatives to thermodynamic stability metrics. These methods treat synthesizability prediction as a classification problem where experimentally reported structures serve as positive examples, while hypothetical structures from computational databases are treated as unlabeled (rather than definitively negative) examples. This paradigm acknowledges that most unsynthesized materials are not inherently unsynthesizable but simply not yet synthesized.

The SynthNN model exemplifies this approach, utilizing a deep learning architecture that learns an optimal representation of chemical formulas directly from the distribution of previously synthesized materials without requiring assumptions about charge balancing or thermodynamic stability [3]. Remarkably, without any prior chemical knowledge, SynthNN learns the chemical principles of charge-balancing, chemical family relationships, and ionicity, utilizing these principles to generate synthesizability predictions [3]. In a head-to-head material discovery comparison against 20 expert material scientists, SynthNN outperformed all experts, achieving 1.5× higher precision and completing the task five orders of magnitude faster than the best human expert [3].

Large Language Models for Crystal Structure Synthesis Prediction

Recent advances have demonstrated the exceptional capability of Large Language Models (LLMs) in predicting synthesizability by learning from text representations of crystal structures. The Crystal Synthesis Large Language Models (CSLLM) framework utilizes three specialized LLMs to predict the synthesizability of arbitrary 3D crystal structures, possible synthetic methods, and suitable precursors, respectively [7]. To enable LLM processing, crystal structures are converted into a text representation termed "material string" that integrates essential crystal information including space group, lattice parameters, and Wyckoff positions in a compact format [7].

Table 2: Performance of CSLLM Framework Components

CSLLM Component	Prediction Task	Performance	Key Innovation
Synthesizability LLM	Binary classification (synthesizable/non-synthesizable)	98.6% accuracy [7]	Outperforms energy (74.1%) and phonon (82.2%) methods
Method LLM	Synthetic route classification (solid-state/solution)	91.0% accuracy [7]	Predicts appropriate synthesis approach
Precursor LLM	Suitable precursor identification	80.2% success rate [7]	Suggests chemical precursors for synthesis

The exceptional performance of LLM-based approaches stems from their ability to learn complex patterns from comprehensive datasets of known materials. These models leverage not only structural and compositional features but also implicit knowledge about synthetic accessibility learned from the entire corpus of reported inorganic crystals. Furthermore, fine-tuned LLMs provide explainable synthesizability predictions, generating human-readable explanations for the factors governing synthesizability and extracting underlying physical rules [2].

Experimental Protocols and Methodologies

Dataset Construction for Synthesizability Prediction

Robust synthesizability prediction requires carefully curated datasets with balanced positive and negative examples. The following protocol outlines the dataset construction process used in state-of-the-art models:

• Positive Example Selection: Extract experimentally confirmed crystal structures from the Inorganic Crystal Structure Database (ICSD). Apply filtering to include only ordered structures with limited elemental diversity (e.g., ≤7 different elements) and reasonable unit cell size (e.g., ≤40 atoms) to ensure tractability [7].
• Negative Example Generation: Calculate CLscore (a synthesizability metric) for a large pool of theoretical structures from materials databases (Materials Project, Computational Materials Database, Open Quantum Materials Database, JARVIS) using pre-trained Positive-Unlabeled learning models. Select structures with the lowest CLscores (e.g., <0.1) as negative examples of non-synthesizable materials [7].
• Composition Balancing: Ensure balanced representation across different crystal systems (cubic, hexagonal, tetragonal, orthorhombic, monoclinic, triclinic, and trigonal) and element combinations to prevent algorithmic bias [7].
• Text Representation: Convert crystal structures to text format using either the "material string" representation (integrating space group, lattice parameters, and Wyckoff positions) or tools like Robocrystallographer that generate natural language descriptions of crystal structures [7] [2].

LLM Fine-Tuning Protocol

For LLM-based synthesizability prediction, the following fine-tuning protocol has proven effective:

• Model Selection: Utilize foundation models like GPT-4o-mini as base architectures, which have demonstrated superior performance compared to earlier models like GPT-3.5 [2].
• Input Formatting: Structure input prompts to include both stoichiometric information and structural descriptions. Experiments show that models incorporating structural information (StructGPT) outperform composition-only models (StoiGPT) for polymorph-specific synthesizability assessment [2].
• Fine-Tuning Strategy: Employ supervised fine-tuning on the curated dataset of positive and negative synthesizability examples. Use appropriate batch sizes and learning rates to maintain the model's general knowledge while adapting it to the specific synthesizability prediction task.
• Embedding Alternative: For cost-efficient deployment, consider extracting embedding representations from the LLM and training separate classifier networks, which can reduce inference costs by 57% compared to full LLM inference [2].

Validation and Benchmarking

Rigorous validation of synthesizability models requires specialized protocols:

• Hold-Out Testing: Reserve 20% of positive and unlabeled data as a hold-out test dataset for assessing model performance [2].
• PU-Learning Metrics: Employ PU-learning specific evaluation metrics including true positive rate (recall) and α-estimation for approximating precision and false positive rates, acknowledging the absence of true negative examples [2].
• Comparative Benchmarking: Compare performance against multiple baselines including random guessing, charge-balancing approaches, formation energy thresholds, and human expert performance [3].
• Generalization Testing: Validate model performance on structurally complex materials with large unit cells that exceed the complexity of training data to assess generalization capability [7].

Essential Research Reagents and Computational Tools

Table 3: Essential Research Reagents and Tools for Synthesizability Research

Resource/Tool	Type	Function/Purpose	Access
ICSD (Inorganic Crystal Structure Database)	Database	Primary source of experimentally confirmed crystal structures for positive examples	Commercial
Materials Project	Database	Source of hypothetical structures for negative example generation	Public
Robocrystallographer	Software Tool	Generates text-based descriptions of crystal structures for LLM input	Open Source
CLscore Model	Computational Model	Generates synthesizability scores for theoretical structures	Research Implementation
Material String Representation	Data Format	Compact text representation of crystal structure for LLM processing	Research Implementation
PU-CGCNN	Computational Model	Graph neural network for synthesizability prediction; baseline comparator	Research Implementation
VASP	Software Package	DFT calculations for traditional thermodynamic stability assessment	Commercial
Fine-Tuned LLMs (e.g., StructGPT)	Computational Model	LLM specialized for synthesizability prediction	Research Implementation

The limitations of thermodynamic stability as a synthesizability proxy are both quantitative and fundamental. With accuracy rates of approximately 74-82% compared to 92-99% for advanced machine learning approaches, thermodynamic metrics alone are insufficient for reliable synthesizability assessment in computational materials discovery. The emerging paradigm of data-driven synthesizability prediction, particularly through LLMs and specialized machine learning models, represents a transformative advancement that directly addresses the complex interplay of thermodynamic, kinetic, and experimental factors governing materials synthesis.

The implications for materials research are profound. By integrating accurate synthesizability predictors into computational screening workflows, researchers can prioritize experimental efforts on genuinely accessible materials with high potential for successful synthesis. Furthermore, the ability to predict not just synthesizability but also appropriate synthetic methods and potential precursors represents a crucial step toward autonomous materials discovery pipelines. As these models continue to evolve, incorporating richer experimental data and more sophisticated representations of synthetic pathways, they will increasingly narrow the gap between computational prediction and experimental realization, accelerating the discovery of novel functional materials for energy, electronics, and beyond.

Within the high-throughput computational search for novel synthesizable inorganic crystals, the principle of charge-balancing stands as a foundational heuristic. This whitepaper examines its role as a critical, yet ultimately limited, filter in materials discovery pipelines. While charge neutrality is a non-negotiable requirement for stable crystalline compounds, an over-reliance on this single metric constitutes the "Charge-Balancing Fallacy"—the assumption that charge balance alone is a sufficient predictor of synthesizability and thermodynamic stability. Through a critical analysis of contemporary literature and emerging benchmarking frameworks, this article delineates the boundaries of charge-balancing's utility. It argues that for computational materials science to overcome its major discovery challenges, it must move beyond this and other isolated heuristics and adopt integrated, quantitative, and probabilistic assessment models that account for the complex multi-dimensional parameter space governing crystal formation.

The discovery of novel inorganic crystalline materials is a cornerstone of technological advancement, underpinning progress in domains from clean energy to biomedicine [8]. However, the traditional experimental discovery process is slow, often averaging two decades from initial research to commercialization, and is inherently limited in its ability to explore the vastness of chemical space [8]. This space is astronomically large; for just quaternary materials, conservative estimates suggest over 10^10 compositions are possible when considering only electronegativity and charge-balancing rules [9].

In response, computational materials science has emerged as a powerful tool for accelerating discovery. The paradigm has shifted from post-rationalizing experimental observations to truly predictive workflows where theory leads experimentation [8]. Central to these high-throughput computational pipelines are filters—computational expressions of human domain knowledge and scientific principles used to screen millions of hypothetical candidate compounds and weed out those that are likely unsynthesizable or unstable [10]. These filters can be categorized as "hard" or "soft":

• Hard Filters encode non-negotiable physical laws. A prime example is charge neutrality, a fundamental requirement for any stable crystalline compound.
• Soft Filters encode empirical rules of thumb, such as the Hume-Rothery rules for solid solubility, which are frequently broken but still provide useful guidance [10].

While indispensable, an over-reliance on any single heuristic, including the foundational principle of charge balance, can become a fallacy that limits discovery. This article examines the precise nature of this Charge-Balancing Fallacy and explores the path toward more robust, integrated discovery frameworks.

The Charge-Balancing Heuristic: Utility and Limitations

The Fundamental Necessity of Charge Balance

The charge-balancing heuristic is rooted in the unequivocal requirement that a stable crystalline compound must be electrically neutral overall. In the context of inorganic crystals, this typically involves ensuring that the positive charges from cations balance the negative charges from anions within a given composition. This principle is so fundamental that it is often one of the first and most strictly applied filters in a materials screening pipeline. Its application drastically reduces the combinatorial search space, making computational surveys of hypothetical materials tractable [10] [9].

The Fallacy: Sufficiency versus Necessity

The "Charge-Balancing Fallacy" arises not from a misunderstanding of its necessity, but from the incorrect assumption of its sufficiency. A charge-balanced composition is a necessary condition for stability, but it is far from a sufficient predictor of actual synthesizability or thermodynamic persistence. This fallacy manifests in several critical ways:

• Neglect of Structural Stability: A composition can be charge-balanced yet correspond to numerous potential atomic arrangements (polymorphs), most of which may be energetically unfavorable. The accurate prediction of the most stable crystal structure—the ground-state configuration—remains one of the most significant challenges in computational materials science [8] [11]. The solid-state packing arrangement is a key driver of a material's properties, and small changes can drastically alter its stability and functionality [8].

• Oversimplification of Thermodynamics: Thermodynamic stability is not determined by a compound's formation energy in isolation, but by its energetic competition with all other phases in its chemical space, quantified by its energy above the convex hull (Ehull) [9]. A charge-balanced compound can easily have a positive Ehull, indicating that it is metastable and will tend to decompose into a mixture of more stable compounds.

• Exclusion of Kinetic and Synthetic Factors: Synthesizability is influenced by kinetic barriers, reaction pathways, and processing conditions, which are not captured by a simple charge-balancing check. A charge-balanced compound may be thermodynamically stable but impossible to synthesize under practical conditions, or it may form undesirable, metastable polymorphs [12].

The following table summarizes key heuristics beyond charge balance that are critical for assessing synthesizability and stability.

Table 1: Key Heuristics and Metrics for Predicting Crystal Stability and Synthesizability

Heuristic/Metric	Type	Description	Limitations
Charge Neutrality [10]	Hard Filter	Ensures the compound's overall charge is zero.	Necessary but insufficient; does not guarantee stability.
Energy Above Hull (E_hull) [9]	Quantitative Metric	Energy difference between a compound and the most stable mixture of other phases from the convex hull phase diagram.	Primary indicator of thermodynamic stability; requires accurate energy calculations.
Structure Prediction Accuracy [11]	Quantitative Challenge	The ability to computationally predict the correct ground-state crystal structure from a composition.	Computationally expensive; accuracy is tied to the method's cost.
Electronegativity Balance [10]	Soft Filter	Checks for reasonable electronegativity differences between elements.	An empirical rule that is frequently violated in known stable compounds.

Quantitative Frameworks for Moving Beyond the Fallacy

To overcome the limitations of isolated heuristics, the field is moving toward standardized, quantitative evaluation frameworks that integrate multiple stability criteria.

Benchmarking Crystal Structure Prediction

The performance of Crystal Structure Prediction (CSP) algorithms is critical, as accurate structure prediction is a prerequisite for reliable property calculation. Historically, evaluating CSP algorithms relied heavily on manual inspection and comparison of formation energies [11]. This lack of standardized metrics made it difficult to compare different methods objectively.

Recent work has focused on developing quantitative CSP performance metrics that automatically determine the quality of predicted structures against known ground truths. These include a suite of structure similarity metrics that, when combined, capture key aspects of structural fidelity, even when predicted structures have different spatial symmetries than the target [11]. The move toward such automated, quantitative benchmarking is essential for rigorously evaluating and improving the computational tools that underpin modern materials discovery.

Prospective vs. Retrospective Evaluation

A major disconnect in the field has been between retrospective benchmarking on known, stable materials and prospective performance in a genuine discovery campaign targeting unknown materials. To address this, frameworks like Matbench Discovery have been developed to simulate real-world discovery [9].

This framework highlights a critical misalignment: a model can exhibit excellent regression performance (e.g., low Mean Absolute Error in formation energy) but still have a high false-positive rate for stable materials if its predictions lie close to the stability decision boundary (0 eV/atom Ehull). This underscores why evaluation must be based on task-relevant classification metrics (e.g., precision and recall for stability) rather than regression accuracy alone [9]. The following table contrasts different model evaluation approaches.

Table 2: Comparison of Model Evaluation Paradigms in Materials Discovery

Evaluation Aspect	Traditional/Restricted Approach	Advanced/Prospective Approach	Implication
Primary Metric	Regression accuracy (e.g., MAE of formation energy) [9]	Classification performance (e.g., F1-score for stability) [9]	Focuses on correct decision-making, not just numerical accuracy.
Data Splitting	Random split on known materials [9]	Time-split or cluster-based split to test generalizability [9]	Better simulates the challenge of finding truly novel materials outside the training distribution.
Stability Target	Formation energy [9]	Energy above the convex hull (Ehull) [9]	Directly measures thermodynamic stability against phase decomposition.
Structure Handling	Using relaxed DFT structures as input [9]	Predicting from unrelaxed (initial) structures [9]	Avoids circular logic and increases practical utility for screening new candidates.

Experimental Protocols and the Scientist's Toolkit

Integrated Workflow for Stable Crystal Discovery

A modern, robust pipeline for discovering synthesizable inorganic crystals integrates high-throughput computation, machine learning, and high-fidelity validation. The following diagram visualizes this multi-stage workflow, highlighting how heuristics like charge balancing are embedded within a broader, more rigorous process.

Workflow for Discovering Stable Crystals

The Scientist's Computational Toolkit

The following table details key computational "reagents" and resources essential for executing the workflow described above.

Table 3: Essential Computational Tools for Predicting Synthesizable Crystals

Tool/Resource	Category	Function	Example Tools
Hypothetical Databases	Data	Large datasets of enumerated hypothetical compounds for screening.	Synthetic datasets from various sources [10]
CSP Algorithms	Software	Predicts the stable crystal structure of a given composition.	USPEX, CALYPSO, AIRSS [11]
Universal Interatomic Potentials (UIPs)	Model	Machine learning force fields for fast, accurate energy and force predictions.	UIPs highlighted in Matbench Discovery [9]
First-Principles Methods	Software	High-fidelity quantum mechanical calculations for validation.	Density Functional Theory (DFT) with VASP [11] [9]
Stability Metrics	Metric	Quantifies thermodynamic stability.	Energy above convex hull (E_hull) [9]
Benchmarking Platforms	Framework	Standardized evaluation of ML and CSP algorithm performance.	Matbench Discovery [9], CSPBenchMetrics [11]

Free-Energy Calculation and Error Quantification Protocol

For the highest level of confidence, particularly in industrial applications like pharmaceutical polymorph selection, advanced free-energy calculation protocols have been developed. A state-of-the-art method, as demonstrated in recent studies, involves a composite approach [12]:

• Composite Energy Calculation: Combine a hybrid density functional (PBE0) with many-body dispersion (MBD) corrections and vibrational free energy (Fvib) contributions to achieve high accuracy.
• Error Quantification: Critically, establish a reliable experimental benchmark for solid-solid free-energy differences. From this, derive transferable error estimates, typically expressed as a standard error per atom (e.g., ~0.191 kJ mol⁻¹) and per water molecule for hydrates (~0.641 kJ mol⁻¹) [12].
• Phase Diagram Construction: Use the computed free energies with their associated errors to place anhydrate and hydrate crystal structures on the same energy landscape as a function of temperature and relative humidity, enabling predictive risk assessment of phase transitions.

This methodology transforms CSP from a qualitative tool into a quantitative, actionable procedure where predictions come with defined error bars, allowing for robust decision-making [12].

The charge-balancing heuristic is a necessary first gatekeeper in the computational search for new inorganic materials, but falling into the "Charge-Balancing Fallacy" by treating it as a sufficient condition for synthesizability severely limits discovery potential. The path forward lies in embracing integrated and probabilistic workflows that synthesize multiple lines of evidence.

The future of materials discovery will be driven by:

• Multi-Filter Pipelines: Embedding charge balance alongside other hard and soft filters (e.g., electronegativity, structural motifs) within a systematic screening environment [10].
• Advanced ML and UIPs: Leveraging universal interatomic potentials and other machine learning models that have been prospectively benchmarked for their ability to accurately classify stability, not just regress formation energies [9].
• Quantified Uncertainty: Adopting frameworks that provide error estimates for computed free energies, enabling true risk assessment and prioritization of experimental targets [12].
• Closing the Loop with Experiment: Using computation to guide synthesis and, in turn, using experimental results to refine and validate computational models prospectively.

By moving beyond the charge-balancing fallacy and other isolated heuristics, the field can fully exploit the enormous potential materials space and systematically discover the novel functional materials urgently needed to address global challenges.

The discovery of novel inorganic crystalline materials is a fundamental driver of technological progress, from developing more efficient batteries to designing new pharmaceuticals. While computational models can rapidly generate millions of hypothetical crystal structures with desirable properties, a critical bottleneck persists: the overwhelming majority of these virtual candidates cannot be synthesized in a laboratory [13] [3]. This discrepancy creates a significant gap between theoretical prediction and experimental realization, slowing the entire discovery cycle. The core of this problem is a fundamental data scarcity issue. In a typical supervised machine learning (ML) classification task, a model is trained on a balanced set of both positive examples (synthesizable crystals) and negative examples (non-synthesizable crystals). However, in materials science, while vast databases of successfully synthesized materials exist (e.g., the Materials Project (MP) and the Inorganic Crystal Structure Database (ICSD)), there are no definitive repositories of "unsynthesizable" materials [3] [2]. Failed synthesis attempts are rarely reported in the scientific literature, and the space of impossible compounds is astronomically large and undefined [14] [3]. This lack of verified negative data renders standard binary classification ML models ineffective for predicting synthesizability. To overcome this "data hurdle," researchers have turned to Positive-Unlabeled (PU) Learning, a class of semi-supervised machine learning techniques designed to learn exclusively from positive and unlabeled data [2] [15]. This whitepaper provides an in-depth technical guide to the PU learning problem, its methodologies, and its application as an essential framework for predicting the synthesizability of inorganic crystals.

The PU Learning Paradigm: Core Concepts and Formulations

Problem Formulation and Key Assumptions

Positive-Unlabeled (PU) learning formalizes the synthesizability prediction problem by redefining the available data. The set of all experimentally synthesized crystals, typically sourced from the ICSD or MP, is treated as the Positive (P) set. The vast space of hypothetical, computer-generated crystals for which synthesis has not been attempted or confirmed is treated not as negative, but as the Unlabeled (U) set. The key insight is that the unlabeled set is a mixture of both actually positive (synthesizable but yet-to-be-made) and actually negative (unsynthesizable) examples; the learner's task is to identify the hidden negative examples within this unlabeled set [3] [2]. The success of PU learning relies on two foundational assumptions:

• Selected Completely at Random (SCAR) Assumption: The probability that a positive example is labeled (i.e., included in the known synthesized database) is constant and independent of its attributes [3].
• Positive Examples are "Typical": The known positive examples in the labeled set are representative of the entire population of synthesizable materials. If the labeled positives are a biased subset (e.g., only containing certain crystal systems), the model's ability to generalize will be compromised.

A Comparative View of Traditional and PU-Based Approaches

The table below contrasts traditional stability metrics and standard ML with the PU learning approach, highlighting why PU learning is necessary for this domain.

Table 1: Comparison of Synthesizability Assessment Methods

Method Category	Core Principle	Key Advantage	Fundamental Limitation
Thermodynamic Stability	Uses Energy Above Hull (E$_{\text{hull}}$) as a proxy for stability.	Physically intuitive; easily calculated via Density Functional Theory (DFT).	Fails to account for kinetics and synthesis conditions; many metastable materials exist [14] [7].
Charge Balancing	Filters compositions based on net-neutral ionic charge.	Computationally inexpensive; chemically motivated.	Inflexible; fails for metallic/covalent materials; poor empirical accuracy (e.g., only 37% of ICSD materials are charge-balanced) [3].
Standard Supervised ML	Trains a binary classifier on known positive and negative examples.	Powerful if representative negative examples are available.	Inapplicable due to the complete lack of true negative data [3] [2].
Positive-Unlabeled (PU) Learning	Learns from synthesized (P) and hypothetical (U) materials, treating U as a mixture.	Directly addresses the core data scarcity problem; does not require negative examples.	Performance evaluation is challenging; relies on the SCAR assumption, which may not always hold perfectly [13] [2].

Contemporary Methodologies and Experimental Protocols

Recent research has produced several sophisticated PU learning frameworks for synthesizability prediction. The table below summarizes the architecture, input data, and key performance metrics of several state-of-the-art models.

Table 2: Summary of Contemporary PU Learning Models for Synthesizability

Model Name	Architecture & Input	Key Innovation	Reported Performance
CPUL (Contrastive Positive Unlabeled Learning) [13]	Two-stage model: 1) Contrastive learning for feature extraction, 2) MLP classifier with PU learning.	Uses contrastive learning to pre-train features, improving robustness and reducing PU training time.	True Positive Rate (TPR): 93.95% on MP test set; 88.89% on Fe-containing materials.
SynthNN [3]	Deep learning model using atom2vec composition embeddings.	Learns synthesizability directly from the distribution of all known compositions; requires no crystal structure.	7x higher precision than E$_{\text{hull}}$; outperformed human experts in discovery tasks.
CSLLM (Synthesizability LLM) [7]	Fine-tuned Large Language Model (LLM) using a novel "material string" text representation of crystal structure.	Achieves high accuracy by leveraging LLMs' pattern recognition on a balanced, structure-based dataset.	Accuracy: 98.6% on testing data, significantly outperforming E$_{\text{hull}}$ (74.1%) and phonon stability (82.2%).
PU-GPT-Embedding [2]	Pipeline: 1) Text description of crystal structure → 2) GPT text embeddings → 3) Neural network PU classifier.	Combines the rich representation of LLM embeddings with a dedicated PU classifier, offering high performance and cost efficiency.	Outperforms both graph-based models (PU-CGCNN) and fine-tuned LLMs (StructGPT) in prediction quality.

Detailed Experimental Protocol: A Representative Workflow

The following protocol outlines the key steps for developing and validating a PU learning model for crystal synthesizability, synthesizing common elements from the cited research.

• Data Curation and Preprocessing:

  • Positive Set (P): Extract crystal structures and their compositions from a database of experimentally synthesized materials, such as the ICSD or the synthesized entries in the Materials Project (MP). A typical dataset may include ~38,000 positive samples [13] [2].
  • Unlabeled Set (U): Compile a set of hypothetical crystal structures from generative models or stability screenings (e.g., from MP, OQMD, JARVIS). This set is typically much larger, often containing over 100,000 entries [13] [7].
  • Feature Extraction: Convert the raw crystal data into a numerical representation suitable for ML.
    • For Composition-based Models (e.g., SynthNN): Use learned composition embeddings like atom2vec [3].
    • For Structure-based Models (e.g., CPUL, CSLLM): Generate features using crystal graphs (CGCNN), contrastive learning, or text-based representations like "material strings" or Robocrystallographer descriptions [13] [7] [2].

• Model Training with PU Learning:

  • Algorithm Selection: Choose a PU learning algorithm. A common and effective choice is a "bagging SVM" approach, which involves:
    • Randomly sub-sampling potential negative examples from the unlabeled set.
    • Training an ensemble of classifiers (e.g., Support Vector Machines) on the positive set and each sub-sample.
    • Iteratively re-weighting the unlabeled samples and averaging the scores from all classifiers to produce a final Crystal-Likeness Score (CLscore) between 0 and 1 [13] [15].
  • Loss Function: The model is trained with a loss function that penalizes misclassification of known positive examples while cautiously learning from the unlabeled set. The cost-sensitive binary cross-entropy loss is often used for neural network-based PU learners [3] [2].

• Validation and Performance Assessment:

  • Hold-out Test: Reserve a portion of the known positive data (e.g., 10,000 materials) as a test set. The model's True Positive Rate (TPR) or recall on this set is a primary metric, as it measures the ability to correctly identify synthesizable materials [13] [2].
  • α-Estimation: Because true negatives are unavailable, precision and false positive rates cannot be directly calculated. Methods like α-estimation are used to approximate these metrics by estimating the fraction of positive examples in the unlabeled set [2].
  • Case Study Validation: Apply the trained model to a specific, challenging subset of materials (e.g., all Fe-containing compounds or a novel chemical system) to demonstrate generalizability beyond the training distribution [13] [14].

Visualization of Methodologies and Workflows

High-Level Workflow for Synthesizability Prediction

The following diagram illustrates the end-to-end workflow for building and applying a PU learning model for synthesizability prediction.

Architecture of a Contrastive PU Learning (CPUL) Model

This diagram details the two-stage architecture of a specific advanced model, CPUL, which combines contrastive learning with PU learning.

Table 3: Essential Research Reagents and Computational Tools

Item / Resource	Function / Description	Relevance to PU Learning Experiments
Materials Project (MP) Database [13] [2]	A repository of computed and experimentally known crystal structures and properties.	Primary source for both positive (synthesized) and unlabeled (hypothetical) data.
Inorganic Crystal Structure Database (ICSD) [3] [7]	The world's largest database of fully determined inorganic crystal structures.	The definitive source for high-quality, curated positive examples.
Python Materials Genomics (pymatgen) [13]	A robust, open-source Python library for materials analysis.	Used for parsing CIF/POSCAR files, manipulating crystal structures, and computing features.
Robocrystallographer [2]	A tool that generates text descriptions of crystal structures from CIF files.	Converts structural data into human-readable text for fine-tuning LLMs or creating embeddings.
Crystal-Likeness Score (CLscore) [13] [7]	A probabilistic score (0-1) representing a model's confidence that a material is synthesizable.	The key output metric for ranking and screening candidate materials.
Bagging SVM / Iterative Classifier	A specific PU learning algorithm that ensembles multiple classifiers.	The core engine for many PU models, enabling robust learning from unlabeled data [13] [15].

The integration of Positive-Unlabeled learning has fundamentally shifted the paradigm of synthesizability prediction in computational materials science. By directly confronting the "data hurdle" of negative example scarcity, PU learning provides a principled and effective framework for prioritizing hypothetical crystals for experimental synthesis. The field is rapidly advancing, with current research trends focusing on enhancing model explainability, integrating multimodal data (e.g., synthesis recipes from text-mined literature), and leveraging the power of large foundation models [14] [7] [2]. The development of accurate synthesizability predictors is no longer a distant goal but an active research area, poised to dramatically accelerate the design and discovery of the next generation of functional materials.

The discovery of novel functional materials is a primary driver of technological innovation across fields ranging from energy storage to electronics. A persistent challenge in computational materials science is the significant gap between theoretically predicted materials and those that can be experimentally realized in the laboratory. This challenge necessitates a precise framework for categorizing materials based on their synthesis status and potential. Within the context of predicting synthesizable inorganic crystals, we define three critical classifications: synthesized materials (those experimentally realized and reported in literature), synthesizable materials (those theoretically predicted to be synthetically accessible but not yet synthesized), and unsynthesized materials (a broader category including both synthesizable and fundamentally unsynthesizable compounds). The core research problem lies in accurately distinguishing synthesizable candidates from the vast chemical space of unsynthesized materials, thereby bridging the divide between computational prediction and experimental realization.

Defining the Synthesizability Landscape

Conceptual Framework and Terminology

The relationship between synthesized, synthesizable, and unsynthesized materials can be visualized as a series of intersecting and non-intersecting sets within the total chemical space, each defined by specific criteria related to experimental realization and theoretical potential.

As illustrated in Figure 1, the relationship between these categories is dynamic and evolutionary. The synthesizable set contains materials that meet specific computational or theoretical criteria indicating synthetic accessibility, while the synthesized set represents the subset that has been experimentally realized. Materials may transition from synthesizable to synthesized through experimental effort, while theoretical advances may reclassify certain unsynthesizable materials as synthesizable.

Quantifying the Known and Unknown

Table 1: Representative Data Sources for Synthesized and Hypothetical Materials

Database/Resource	Content Type	Size/Scale	Primary Use in Synthesizability Research
Inorganic Crystal Structure Database (ICSD) [3] [16]	Experimentally synthesized inorganic crystals	~70,120 curated structures (example dataset) [17]	Primary source of positive examples (synthesized materials)
Materials Project (MP) [16] [13]	DFT-calculated structures (mixed synthesized and hypothetical)	>126,000 materials [16]	Source of both positive and unlabeled examples
MatSyn25 (2D Materials) [18]	Synthesis process information from literature	163,240 synthesis processes from 85,160 articles [18]	Training data for synthesis route prediction
OQMD/AFLOW [17]	High-throughput computational data	Millions of calculated structures [17]	Source of hypothetical/unlabeled materials

Computational Approaches for Synthesizability Prediction

Machine Learning Methodologies

Positive-Unlabeled Learning Frameworks

A significant challenge in training synthesizability prediction models is the lack of definitive negative examples (truly unsynthesizable materials). Positive-unlabeled learning addresses this by treating unsynthesized materials as unlabeled rather than negative examples.

SynthNN Implementation: This deep learning model leverages atom2vec representations to learn optimal chemical formula representations directly from data without prior chemical knowledge. Remarkably, it learns chemical principles like charge-balancing and ionicity without explicit programming [3]. The model architecture uses a semi-supervised approach that probabilistically reweights unlabeled examples according to their likelihood of being synthesizable.

Contrastive Positive-Unlabeled Learning (CPUL): This hybrid approach combines contrastive learning with PU learning to predict crystal-likeness scores (CLscore). The framework first employs crystal graph contrastive learning to extract structural and synthetic features, followed by a multilayer perceptron classifier with PU learning to predict CLscore [13]. This approach achieves a true positive rate of 88.89% on Fe-containing materials from the Materials Project database [13].

Large Language Models for Synthesizability Assessment

The Crystal Synthesis Large Language Models framework represents a breakthrough in synthesizability prediction, utilizing three specialized LLMs to predict synthesizability, synthetic methods, and suitable precursors respectively [17]. Key innovations include:

• Material String Representation: A text-based representation of crystal structures that integrates essential crystal information for efficient LLM processing
• Balanced Dataset Curation: 70,120 synthesizable structures from ICSD paired with 80,000 non-synthesizable structures identified through CLscore thresholding
• Domain-Focused Fine-Tuning: Alignment of linguistic features with material features critical to synthesizability

The Synthesizability LLM achieves 98.6% accuracy, significantly outperforming traditional thermodynamic (74.1%) and kinetic (82.2%) stability metrics [17].

Network Science Approaches

The materials stability network approach constructs a scale-free network from the convex free-energy surface of inorganic materials combined with experimental discovery timelines [19]. This network evolves over time with a degree distribution following a power-law (p(k) ~ k^(-γ) with γ = 2.6 ± 0.1 after the 1980s [19].

Key network properties used for prediction include:

• Degree and eigenvector centralities
• Mean shortest path length
• Mean degree of neighbors
• Clustering coefficient

This approach implicitly captures circumstantial factors beyond thermodynamics that influence discovery, including scientific and non-scientific effects such as availability of kinetically favorable pathways, development of new synthesis techniques, and shifts in research focus [19].

Table 2: Performance Comparison of Synthesizability Prediction Methods

Method/Model	Approach Type	Key Metrics	Advantages	Limitations
SynthNN [3]	Deep Learning (Atom2Vec)	7× higher precision than formation energy; Outperforms human experts	Requires no prior chemical knowledge; Learns optimal descriptors from data	Composition-based only (no structure)
FTCP + DL [16]	Fourier-Transformed Crystal Properties	82.6% precision, 80.6% recall for ternary crystals	Incorporates both real and reciprocal space information	Requires structural information
CPUL [13]	Contrastive + PU Learning	93.95% TP accuracy on MP test set	Short training time; High accuracy on limited knowledge	Two-stage process more complex
CSLLM [17]	Large Language Model	98.6% synthesizability accuracy; >90% method/precursor accuracy	Highest accuracy; Predicts methods and precursors	Requires extensive fine-tuning
Stability Network [19]	Network Science	Captures discovery dynamics	Incorporates historical discovery patterns	Indirect synthesizability assessment

Experimental Validation and Synthesis Planning

Retrosynthesis Prediction for Inorganic Solids

Predicting synthesis pathways represents a critical step beyond binary synthesizability classification. The ElemwiseRetro model formulates inorganic retrosynthesis problems by dividing chemical elements into "source elements" (must be provided as precursors) and "non-source elements" (can come from reaction environments) [20].

Element-wise Graph Neural Network Architecture:

• Target composition encoded as a graph with node features from pretrained representations
• Source element mask discriminates source elements from given compositions
• Precursor classifier predicts precursors from a template library
• Joint probability calculation ranks precursor sets

This approach achieves 78.6% top-1 and 96.1% top-5 exact match accuracy, significantly outperforming popularity-based baseline models (50.4% top-1 accuracy) [20]. The probability score strongly correlates with prediction accuracy, providing a confidence metric for experimental prioritization.

Integration of Thermodynamic and Kinetic Factors

While thermodynamic stability (formation energy, energy above convex hull) provides a foundational synthesizability filter, it is insufficient alone. Only 37% of synthesized inorganic materials are charge-balanced according to common oxidation states, and even among typically ionic binary cesium compounds, only 23% are charge-balanced [3]. This demonstrates the limitations of simplistic thermodynamic heuristics.

Successful synthesizability models integrate multiple stability criteria:

• Energy above hull (Ehull) thresholds (e.g., <0.08 eV/atom) [16]
• Phonon stability (absence of imaginary frequencies)
• Phase competition metrics
• Decomposition pathway analysis

Essential Research Tools and Protocols

Table 3: Essential Research Resources for Synthesizability Prediction

Resource/Reagent	Type	Function/Role	Example Applications
ICSD Database [3] [17]	Experimental Database	Primary source of synthesized material structures; Ground truth for model training	Positive examples for supervised learning; Benchmarking
Materials Project API [16] [13]	Computational Database	Access to DFT-calculated properties and structures	Feature extraction; Training data generation
PyMatGen [16] [13]	Python Library	Materials analysis and processing	Structure manipulation; Feature generation
CrabNet [16]	Attention-based Network	Compositional property prediction	Baseline model comparison
CGCNN [16] [13]	Graph Neural Network	Structure-property prediction	Crystal representation learning
FTCP Representation [16]	Crystal Representation	Encodes periodicity and elemental properties	Input for deep learning models

Experimental Workflow for Synthesizability Assessment

The precise definition of the synthesizable space represents a critical advancement in materials discovery, addressing the fundamental challenge of bridging computational prediction and experimental realization. The development of sophisticated machine learning approaches—from positive-unlabeled learning to large language models—has dramatically improved our ability to distinguish synthesizable materials within the vast chemical space of unsynthesized compounds. These computational tools, integrated with experimental validation workflows, provide researchers with a systematic framework for prioritizing synthesis efforts.

Future advancements will likely focus on several key areas: (1) improved integration of kinetic and processing factors into synthesizability predictions, (2) development of standardized material representations for more effective knowledge transfer across domains, and (3) creation of larger, more comprehensive synthesis databases to support data-driven approaches. As these methodologies mature, the systematic identification of synthesizable materials will accelerate the discovery and deployment of novel functional materials, ultimately reducing the time from conceptual design to practical implementation.

AI-Driven Solutions: From Compositional Models to Precursor Prediction

The discovery of new inorganic crystalline materials is fundamental to technological advancement, yet a significant bottleneck persists: the synthesizability gap. This refers to the challenge of predicting whether a computationally designed material can be successfully synthesized in a laboratory. Traditional proxies for synthesizability, such as thermodynamic stability calculated via Density Functional Theory (DFT) or simple charge-balancing rules, often prove inadequate as they fail to capture the complex kinetic and chemical factors governing real-world synthesis [3]. This whitepaper explores the fundamental challenges in predicting synthesizable inorganic crystals and details how deep learning models, particularly synthesizability classification models like SynthNN (Synthesizability Neural Network), are addressing this core problem. We provide an in-depth examination of SynthNN's architecture, training methodology, and performance, while also situating it within the broader landscape of emerging deep learning approaches, including large language models (LLMs) that are pushing the boundaries of accuracy and explainability in synthesizability prediction [3] [17] [2].

The journey of materials discovery has evolved through several paradigms, from empirical trial-and-error to computational simulation and now into a data-driven era. High-throughput computational screening and generative models can propose millions of candidate materials with desirable properties [21] [22]. However, the vast majority of these theoretically "stable" candidates may not be synthetically accessible, creating a critical bottleneck. The central challenge lies in the fact that synthesizability is a multifaceted concept influenced by:

• Thermodynamic and Kinetic Factors: While a negative formation energy is often a prerequisite, it is not sufficient. Kinetic barriers, nucleation rates, and finite-temperature effects play decisive roles [22].
• Chemical Intuition and Rules: Heuristics like charge-balancing, while chemically motivated, are inflexible. Remarkably, only about 37% of known synthesized inorganic materials in the Inorganic Crystal Structure Database (ICSD) are charge-balanced according to common oxidation states, highlighting the limitation of this approach [3].
• Non-Physical Constraints: Practical synthesizability is also governed by external factors such as precursor cost, equipment availability, and human-perceived importance of the target material [3].

This complex interplay of factors makes synthesizability prediction an ideal candidate for data-driven machine learning approaches. By learning directly from the vast and growing database of known synthesized materials, deep learning models can internalize the complex, often implicit, "rules" of inorganic synthesis without relying on potentially incomplete human-defined physical principles.

Core Architecture of SynthNN

SynthNN represents a pioneering deep learning framework that reformulates material discovery as a synthesizability classification task. Its architecture is designed to leverage the entire space of known inorganic chemical compositions to make predictions without requiring prior crystal structure information [3] [23].

Model Input and Atom2Vec Representation

A key innovation of SynthNN is its use of the atom2vec representation. Instead of using pre-defined chemical descriptors, SynthNN represents each chemical formula by a learned atom embedding matrix that is optimized alongside all other parameters of the neural network [3].

• Input: A chemical formula (e.g., "NaCl").
• Representation Learning: The model learns a continuous, dense vector representation (embedding) for each element in the periodic table. The dimensionality of this embedding is a tunable hyperparameter.
• Advantage: This approach allows SynthNN to discover the optimal representation of chemical formulas directly from the distribution of synthesized materials, without requiring assumptions about which factors (e.g., electronegativity, ionic radius) are most important for synthesizability. Experiments suggest that through this process, SynthNN autonomously learns fundamental chemical principles such as charge-balancing, chemical family relationships, and ionicity [3].

Neural Network Architecture and Training Protocol

The core of SynthNN is a deep neural network trained using a Positive-Unlabeled (PU) Learning strategy, which is crucial for addressing the inherent data limitations in this field.

• Architecture: The learned atom embeddings are fed into a deep residual neural network (ResNet). The ResNet architecture, with its skip connections, facilitates the training of very deep networks, enabling the model to learn complex, hierarchical features from the compositional data [3] [24].
• PU Learning Formulation:
  • Positive (P) Data: Experimentally synthesized materials are sourced from the Inorganic Crystal Structure Database (ICSD). These are reliable positive examples.
  • Unlabeled (U) Data: A critical challenge is the lack of confirmed "unsynthesizable" materials. SynthNN addresses this by generating a large set of artificial chemical formulas that are not present in the ICSD. These are treated as unlabeled data, as some may be synthesizable but just undiscovered.
  • Training Objective: The model is trained to classify synthesized materials as positive while probabilistically reweighting the unlabeled examples according to their likelihood of being synthesizable. This semi-supervised approach robustly handles the incomplete data labeling [3] [2].
• Hyperparameters: The model tuning involves optimizing key parameters such as the dimensionality of the atom embeddings and the ratio of artificially generated formulas to synthesized formulas used in training (denoted as ( N_{synth} )) [3].

The following diagram illustrates the SynthNN training workflow and architecture.

Performance Benchmarking and Experimental Protocols

The performance of SynthNN has been rigorously benchmarked against both traditional computational methods and human experts, demonstrating its significant advantages.

Quantitative Performance Comparison

SynthNN's performance is quantitatively superior to traditional methods. The table below summarizes its performance against key baselines as reported in the original study [3].

Table 1: Performance comparison of SynthNN against traditional methods for synthesizability classification.

Method	Key Principle	Precision	Key Limitations
SynthNN	Data-driven classification on compositions	7x higher than DFT formation energy	Requires large dataset; "Black-box" nature
DFT Formation Energy	Thermodynamic stability (energy above convex hull)	Baseline (1x)	Captures only ~50% of synthesized materials; Computationally expensive
Charge-Balancing	Net neutral ionic charge from common oxidation states	Lower than SynthNN	Inflexible; Only 37% of known materials are charge-balanced
Random Guessing	Random weighted by class imbalance	Lowest	Not a viable strategy

In a head-to-head material discovery challenge against 20 expert material scientists, SynthNN outperformed all human experts, achieving 1.5x higher precision and completing the task five orders of magnitude faster than the best-performing human [3]. This demonstrates the model's potential to dramatically accelerate the materials discovery cycle.

Experimental Protocol for Model Validation

The standard protocol for training and validating a model like SynthNN involves several key steps, which are detailed in the table below.

Table 2: Experimental protocol for developing and validating a synthesizability classification model.

Stage	Protocol Description	Key Datasets/Tools
1. Data Curation	Extract synthesized inorganic crystalline materials from ICSD. Filter for quality and remove disordered structures.	Inorganic Crystal Structure Database (ICSD) [3] [17]
2. Generating Unlabeled Data	Create a large set of hypothetical chemical formulas not present in ICSD. These serve as the unlabeled set in PU learning.	Combinatorial formula generation; Previous databases of hypothetical materials [3]
3. Data Representation	Convert chemical formulas into a machine-learnable format. No structural information is required.	atom2vec embeddings; Stoichiometric features [3]
4. Model Training (PU Learning)	Train a deep neural network (e.g., ResNet) using a PU learning loss function that distinguishes positive examples from the unlabeled set.	Deep Learning Frameworks (e.g., TensorFlow, PyTorch); Positive-Unlabeled learning algorithms [3] [2]
5. Model Evaluation	Evaluate on a hold-out test set. Use α-estimation to approximate precision and false positive rate due to the lack of true negatives.	Standard ML metrics (Precision, Recall, F1-score); α-estimation for PU learning [2]

The Evolving Landscape: Beyond SynthNN

While SynthNN operates solely on composition, recent advances have expanded the field to include crystal structure-based predictions and the application of large language models (LLMs), leading to substantial gains in accuracy and explainability.

Structure-Aware and Large Language Models

• Crystal Synthesis Large Language Models (CSLLM): This framework utilizes three specialized LLMs, fine-tuned on a comprehensive dataset of 150,120 crystal structures, to predict synthesizability, suggest synthetic methods, and identify suitable precursors. The Synthesizability LLM achieves a state-of-the-art accuracy of 98.6%, significantly outperforming traditional thermodynamic and kinetic stability metrics [17].
• Explainable Synthesizability with LLMs: Fine-tuned LLMs (e.g., GPT-4) can now be used not only for prediction but also for generating human-readable explanations for their synthesizability classifications. This addresses the "black-box" limitation of earlier models like SynthNN. These models can infer reasons based on structural descriptions, such as noting unrealistic coordination environments or unstable polyhedral connections [2].
• Integrated Pipelines: Recent work has combined compositional and structural synthesizability scores in a rank-average ensemble to screen millions of candidate structures. This pipeline successfully identified highly synthesizable candidates, leading to the experimental confirmation of 7 out of 16 target structures in a high-throughput laboratory, including one novel material [22].

The workflow of these advanced, explainable LLM-based approaches is depicted below.

For researchers entering the field of computational synthesizability prediction, the following tools, datasets, and models are essential.

Table 3: Key research reagents and resources for deep learning-based synthesizability prediction.

Resource Name	Type	Function and Utility	Reference/Availability
Inorganic Crystal Structure Database (ICSD)	Database	The primary source of confirmed synthesizable (positive) crystal structures for training and benchmarking.	[3] [17]
Materials Project (MP)	Database	A rich source of DFT-calculated structures, including many hypothetical (unlabeled/theoretical) materials used as negative/unlabeled examples.	[17] [22]
atom2vec	Algorithm/Representation	A representation learning method that converts chemical elements into learnable embedding vectors, capturing implicit chemical relationships.	[3]
Positive-Unlabeled (PU) Learning	Algorithmic Framework	A semi-supervised learning paradigm critical for handling the lack of confirmed negative data in synthesizability prediction.	[3] [2]
Robocrystallographer	Software Tool	Generates text-based descriptions of crystal structures from CIF files, enabling the use of LLMs for structure-based prediction.	[2]
Crystal Diffusion VAE (CDVAE)	Generative Model	A deep learning generative model for crystal structure prediction, often used in conjunction with synthesizability filters for inverse design.	[25]
CSLLM Framework	Predictive Model	A suite of fine-tuned LLMs for end-to-end synthesizability, synthesis method, and precursor prediction.	[17]

The development of deep learning models for synthesizability classification, beginning with composition-based approaches like SynthNN and rapidly advancing towards structure-aware, explainable LLM-based frameworks, represents a paradigm shift in materials discovery. These models directly address the fundamental challenge of the synthesizability gap by learning complex, real-world synthesis constraints from data, moving beyond the limitations of traditional thermodynamic proxies. As these models become more accurate and interpretable, and as they are integrated into end-to-end discovery pipelines—from generative design to experimental synthesis—they hold the promise of drastically accelerating the journey from theoretical prediction to tangible, functional material. The ongoing research in this field, focusing on integrating multiple data modalities and improving model explainability, is crucial for building the reliable, autonomous materials discovery systems of the future.

The discovery of new inorganic crystalline materials is a fundamental driver of innovation in fields ranging from clean energy to electronics. However, a significant bottleneck persists: reliably predicting which hypothetical materials are synthetically accessible. The vastness of chemical space makes exhaustive experimental trial-and-error impractical. Furthermore, unlike organic synthesis, inorganic solid-state chemistry lacks well-understood reaction mechanisms, and synthesizability is influenced by a complex interplay of thermodynamic, kinetic, and human-centric factors [3]. Traditionally, computational screening has relied on proxy metrics like charge-balancing—filtering materials to have a net neutral ionic charge based on common oxidation states. However, this approach is notably inflexible; analysis shows it correctly identifies only 37% of known synthesized inorganic materials and a mere 23% of known binary cesium compounds [3]. This reveals a critical gap in our ability to navigate the chemical space for novel, synthesizable materials. This whitepaper explores how artificial intelligence (AI), by learning directly from the data of known materials, is overcoming these limitations by discerning complex chemical principles like charge-balancing and chemical family relationships, thereby transforming the prediction of synthesizable inorganic crystals.

Quantitative Benchmarks: AI vs. Traditional Methods

The performance of AI models in predicting synthesizability can be quantitatively compared against traditional methods. The following table summarizes key performance metrics from recent studies, highlighting the significant advantage of data-driven approaches.

Table 1: Performance Comparison of Synthesizability Prediction Methods

Method	Core Principle	Key Performance Metric	Value	Reference
Charge-Balancing	Net neutral ionic charge based on common oxidation states	Precision on Known Synthesized Materials	37%	[3]
DFT Formation Energy	Thermodynamic stability with respect to decomposition products	Recall of Synthesized Inorganic Crystalline Materials	~50%	[3]
SynthNN (AI Model)	Deep learning on known material compositions	Precision in Material Discovery	7x higher than DFT	[3]
SynthNN vs. Human Experts	Data-driven synthesizability classification	Precision & Speed	1.5x higher precision; 5 orders of magnitude faster	[3]

Another AI model, GNoME, has demonstrated unprecedented scale in stable crystal structure prediction, identifying 2.2 million new inorganic crystal structures. From this vast set, 380,000 are predicted to be thermodynamically stable, including 528 new lithium-ion conductors critical for advanced battery technology [26]. These quantitative results underscore a paradigm shift from theory-heavy to data-driven discovery.

Experimental Protocols: How AI Models Learn from Data

The ability of AI to learn intricate chemical rules is not pre-programmed but emerges from specific experimental designs and training methodologies. The protocols for key models illustrate this process.

Protocol 1: Deep Learning Synthesizability Classification (SynthNN)

Objective: To train a deep learning model (SynthNN) to classify inorganic chemical formulas as synthesizable or unsynthesizable without requiring structural information [3].

• Data Curation:

  • Positive Data: Extract chemical formulas of synthesized crystalline inorganic materials from the Inorganic Crystal Structure Database (ICSD).
  • Negative Data: Artificially generate a large set of chemical formulas that are not present in the ICSD. These are treated as "unsynthesized" but are acknowledged to potentially include synthesizable materials (Positive-Unlabeled learning).

• Material Representation:

  • Employ an atom2vec framework. This method represents each chemical formula using a learned atom embedding matrix that is optimized alongside all other parameters of the neural network.
  • The model automatically learns an optimal numerical representation of chemical compositions directly from the distribution of the training data, without manual feature engineering.

• Model Training with Positive-Unlabeled Learning:

  • Train a deep neural network on the curated dataset.
  • Implement a semi-supervised learning approach that probabilistically re-weights the artificially generated "unsynthesized" examples according to their likelihood of actually being synthesizable. This accounts for the incomplete labeling in the data.

• Validation:

  • Benchmark SynthNN's performance against random guessing, charge-balancing baselines, and DFT-based formation energy calculations.
  • Conduct a head-to-head comparison where the model and human experts perform the same material discovery task.

Protocol 2: Text-Guided Crystal Structure Generation (Chemeleon)

Objective: To generate chemical compositions and crystal structures by learning from textual descriptions and 3D structural data [27].

• Cross-Modal Contrastive Learning (Crystal CLIP):

  • Text Encoder: A transformer-based model (e.g., MatTPUSciBERT) is pre-trained on a large corpus of materials science literature.
  • Structure Encoder: An Equivariant Graph Neural Network (GNN) generates embedding vectors from crystal structures.
  • Training: The model is trained to maximize the cosine similarity (align closely) between text embeddings and graph embeddings derived from the same crystal structure (positive pairs), while minimizing the similarity for mismatched pairs (negative pairs). This creates a shared latent space where text and structure representations are aligned.

• Generative Diffusion Model:

  • Forward Process: Gradually add random noise to a crystal structure's representation (atom coordinates, lattice vectors) over many steps until it becomes pure noise.
  • Backward (Denoising) Process: Train an equivariant GNN to iteratively predict and remove the noise, reconstructing a crystal structure from noise.
  • Conditioning: Use classifier-free guidance, where the text embeddings from the Crystal CLIP model are fed as conditioning input to the denoising model. This guides the generation process towards structures that match the textual description (e.g., "Zn-Ti-O ternary compound").

Protocol 3: Ranking-Based Retrosynthesis (Retro-Rank-In)

Objective: To recommend and rank sets of precursor materials for synthesizing a target inorganic compound [28].

• Problem Reformulation: Frame retrosynthesis not as a multi-label classification task, but as a pairwise ranking problem. This allows the model to recommend precursors it never encountered during training.

• Model Architecture:

  • Composition Encoder: A transformer-based model generates chemically meaningful vector representations for both target materials and potential precursors.
  • Pairwise Ranker: A separate model is trained to score the chemical compatibility between a target material and a precursor candidate, predicting the likelihood they can co-occur in a viable synthetic route.

• Training: The ranker is trained on a bipartite graph of inorganic compounds, learning to assign high scores to historically reported target-precursor pairs and low scores to incorrect pairs.

Diagram 1: The Retro-Rank-In framework embeds targets and precursors into a shared space via a composition encoder. A pairwise ranker then scores their chemical compatibility.

Visualizing AI Learning Pathways and Workflows

The following diagrams illustrate the core workflows of the AI models discussed, highlighting how they process information to learn chemical principles.

Synthesizability Classification with SynthNN

Diagram 2: SynthNN workflow. The model uses atom embeddings and Positive-Unlabeled learning to classify synthesizability.

Text-Guided Crystal Generation with Chemeleon

Diagram 3: Chemeleon generation. A text prompt is encoded and guides a diffusion model to generate a crystal structure from noise.

The development and application of these AI models rely on a foundation of specific computational tools and datasets. The following table details these essential "research reagents."

Table 2: Key Research Reagents in AI-Driven Materials Discovery

Resource Name	Type	Primary Function in Research
Inorganic Crystal Structure Database (ICSD)	Materials Database	Serves as the primary source of "positive" data (known synthesized materials) for training supervised and semi-supervised models like SynthNN [3].
Materials Project Database	Materials Database	Provides a large repository of computed material properties (e.g., formation energies) used for training generative models like Chemeleon and for incorporating domain knowledge [27] [28].
Graph Neural Networks (GNNs)	Algorithm / Model Architecture	Directly operates on graph representations of molecules and crystals, enabling property prediction and structure generation while respecting physical symmetries [27] [26].
atom2vec / Composition Embeddings	Material Representation	Converts chemical element symbols into continuous-valued vectors, allowing models to learn the periodic trends and chemical similarities directly from data [3].
Denoising Diffusion Models	Generative Algorithm	A state-of-the-art framework for generating high-quality crystal structures by iteratively refining random noise into a coherent structure, often conditioned on text or properties [27].
MatTPUSciBERT / SciBERT	Pre-trained Language Model	Provides a foundational understanding of materials science language, which can be fine-tuned for specific tasks like text-structure alignment in Crystal CLIP [27].

The integration of AI into the prediction of synthesizable inorganic crystals represents a profound shift in materials science methodology. By learning directly from the collective data of known materials, deep learning models internalize complex chemical principles like charge-balancing and chemical family relationships without explicit programming. This data-centric approach has proven to outperform traditional heuristic and thermodynamic-based methods in both precision and scale, as evidenced by models like SynthNN, Chemeleon, and GNoME. Furthermore, the development of explainable AI techniques, such as the Substructure Mask Explanation (SME) method, is beginning to open the "black box," providing chemists with intuitive, fragment-based insights into model predictions [29]. While challenges remain—including the need for standardized benchmarks and further experimental validation—the ability of AI to discern the hidden rules of inorganic chemistry from data is fundamentally accelerating the discovery of new materials for clean energy, electronics, and beyond.

The discovery of new functional materials is often bottlenecked by the challenge of synthesizing computationally predicted candidates. Traditional methods that use thermodynamic or kinetic stability as a proxy for synthesizability exhibit significant limitations, capturing only 50-82% of synthesizable materials. The Crystal Synthesis Large Language Models (CSLLM) framework represents a paradigm shift, leveraging three specialized large language models to directly predict synthesizability, synthesis methods, and suitable precursors for arbitrary 3D crystal structures. CSLLM achieves remarkable accuracy (98.6%) in synthesizability prediction, significantly outperforming traditional approaches and demonstrating exceptional generalization to complex structures. This technical guide examines CSLLM's architecture, methodology, and performance within the broader context of overcoming fundamental challenges in inorganic crystal synthesis prediction.

Computational materials science has advanced to the point where machine learning and high-throughput screening can generate millions of theoretical candidate materials with promising properties. However, a critical gap persists between in silico predictions and real-world laboratory synthesis. This disconnect stems from several fundamental challenges:

• Thermodynamic Limitations: Conventional synthesizability screening relies on density functional theory (DFT) calculations of formation energies or energy above the convex hull (ΔEhull). However, numerous structures with favorable formation energies remain unsynthesized, while various metastable structures are successfully synthesized despite less favorable thermodynamics [7]. Thermodynamic stability alone captures only approximately 74.1% of synthesizable structures [7].

• Kinetic Stability Limitations: Phonon spectrum analysis assesses kinetic stability but is computationally expensive. Moreover, materials with imaginary phonon frequencies can still be synthesized, indicating this metric's limitations [7].

• Charge-Balancing Inadequacy: Charge-balancing based on common oxidation states performs poorly as a synthesizability proxy, correctly identifying only 37% of known synthesized inorganic materials and merely 23% of binary cesium compounds [3].

• Structural Knowledge Dependency: Many machine learning approaches require complete crystal structure information, which is typically unknown for undiscovered materials [3].

The Crystal Synthesis Large Language Models (CSLLM) framework addresses these limitations by learning the complex patterns underlying successful synthesis directly from comprehensive datasets of known materials, enabling more accurate and practical predictions of synthesizability and synthesis pathways.

CSLLM Framework: Architecture and Methodology

The CSLLM framework employs a multi-component architecture comprising three specialized large language models, each fine-tuned for specific aspects of the synthesis prediction problem [7] [30]:

• Synthesizability LLM: Predicts whether an arbitrary 3D crystal structure is synthesizable
• Method LLM: Classifies appropriate synthesis methods (e.g., solid-state or solution)
• Precursor LLM: Identifies suitable chemical precursors for synthesis

This specialized approach allows each model to develop expertise in its respective domain while enabling comprehensive synthesis pathway planning when used together.

Data Curation and Representation

A critical innovation underpinning CSLLM is the construction of a comprehensive, balanced dataset for training and evaluation:

Table 1: CSLLM Dataset Composition

Data Category	Source	Selection Criteria	Count	Purpose
Synthesizable Structures	Inorganic Crystal Structure Database (ICSD)	≤40 atoms, ≤7 elements, no disordered structures	70,120	Positive examples
Non-synthesizable Structures	Materials Project, CMD, OQMD, JARVIS	CLscore <0.1 from PU learning model [7]	80,000	Negative examples
Total	Multiple sources	Comprehensive coverage	150,120	Model training/validation

The dataset encompasses seven crystal systems (cubic, hexagonal, tetragonal, orthorhombic, monoclinic, triclinic, and trigonal) and elements with atomic numbers 1-94 (excluding 85 and 87), ensuring broad chemical diversity [7].

To enable LLM processing, the researchers developed a novel text representation called "material string" that efficiently encodes essential crystal structure information in a compact format. This representation includes space group information, lattice parameters (a, b, c, α, β, γ), and atomic coordinates with Wyckoff positions, eliminating redundancies present in conventional CIF or POSCAR formats [7].

Model Training and Experimental Protocol

The CSLLM framework implementation follows a rigorous experimental protocol:

• Data Preprocessing: Conversion of crystal structures to material string representation, including symmetry analysis and Wyckoff position determination.

• Model Architecture Selection: Utilization of foundation LLMs (architecture unspecified in available literature) as base models for domain-specific fine-tuning.

• Domain Adaptation: Fine-tuning on the curated synthesizability dataset using standard language modeling objectives, enabling the models to align linguistic features with materials science concepts relevant to synthesizability.

• Validation Methodology: Employing hold-out test sets and prospective validation on structures with complexity exceeding training data to assess generalization capability.

The fine-tuning process essentially teaches the models to recognize patterns in the material strings that correlate with successful synthesis, leveraging the broad knowledge base of the underlying LLMs while specializing them for the crystallography domain.

Performance Analysis and Benchmarking

Synthesizability Prediction Accuracy

CSLLM's Synthesizability LLM establishes new state-of-the-art performance in crystal synthesizability prediction:

Table 2: Synthesizability Prediction Performance Comparison

Method	Accuracy	Advantage over CSLLM	Key Limitations
CSLLM Synthesizability LLM	98.6%	Baseline (reference)	Requires crystal structure information
Traditional Thermodynamic (ΔEhull ≥0.1 eV/atom)	74.1%	-24.5% (106.1% lower relative accuracy)	Fails for metastable synthesizable compounds
Kinetic Stability (lowest phonon frequency ≥-0.1 THz)	82.2%	-16.4% (44.5% lower relative accuracy)	Computationally expensive; imaginary frequencies don't preclude synthesis
Teacher-Student Dual Neural Network [7]	92.9%	-5.7%	Architecture-specific limitations
Positive-Unlabeled Learning [7]	87.9%	-10.7%	Moderate accuracy

The Synthesizability LLM demonstrates exceptional generalization capability, achieving 97.9% accuracy on additional testing structures with large unit cells that significantly exceed the complexity of the training data [7].

Synthesis Method and Precursor Prediction

The Method and Precursor LLMs show similarly impressive performance:

• Method LLM: Achieves 91.0% accuracy in classifying appropriate synthesis methods (solid-state vs. solution) for target compounds [7] [30].

• Precursor LLM: Attains 80.2% success rate in identifying appropriate solid-state synthesis precursors for binary and ternary compounds [7]. This performance is notable given the combinatorial challenge of precursor selection.

For comparison, alternative approaches for precursor prediction include ElemwiseRetro, a graph neural network-based model that achieves 78.6-80.4% top-1 exact match accuracy in predicting inorganic synthesis precursors [20]. This model formulates retrosynthesis as a source element identification and precursor template selection problem, demonstrating the viability of multiple approaches to this challenging task.

Prospective Validation and Practical Utility

Beyond retrospective benchmarking, CSLLM was prospectively applied to assess the synthesizability of 105,321 theoretical structures, identifying 45,632 as synthesizable [7]. The framework includes a user-friendly interface for automatic prediction from uploaded crystal structure files, enhancing practical utility for materials researchers.

Alternative Approaches and Contextual Framework

While CSLLM represents a significant advancement, other computational approaches address related challenges in materials discovery:

Table 3: Complementary Computational Approaches in Materials Science

Method	Application Scope	Key Innovation	Performance
SynthNN [3]	Composition-based synthesizability prediction	Atom2Vec embeddings; positive-unlabeled learning	7× higher precision than formation energy
Matbench Discovery [9]	ML energy model evaluation	Prospective benchmarking framework	Identifies best-performing methodologies
SPaDe-CSP [31]	Organic crystal structure prediction	ML-based lattice sampling & neural network potentials	80% success rate (2× random sampling)
ElemwiseRetro [20]	Inorganic retrosynthesis	Element-wise graph neural network with templates	78.6-80.4% top-1 precursor accuracy
Diffraction Fingerprinting [32]	Crystal symmetry classification	Deep learning on diffraction images	Robust to defects (up to 40% atom loss)

These complementary approaches highlight the diverse strategies being employed across the materials informatics landscape, with CSLLM occupying the specialized niche of structure-based synthesis prediction via large language models.

Experimental Workflow and Research Toolkit

CSLLM Experimental Workflow

The following diagram illustrates the comprehensive workflow for crystal synthesis prediction using the CSLLM framework:

Table 4: Research Reagent Solutions for Crystal Synthesis Prediction

Resource Type	Specific Examples	Function in Research	Access Information
Structural Databases	ICSD, Materials Project, CMD, OQMD, JARVIS [7]	Sources of confirmed synthesizable and non-synthesizable structures	Public/restricted access
Descriptor Tools	Material string converter, CIF parser, symmetry analysis	Convert crystal structures to LLM-readable format	Custom implementation
Validation Datasets	Complex structures with large unit cells, prospective candidates [7]	Assess model generalization beyond training data	Research publications
Benchmarking Frameworks	Matbench Discovery [9]	Standardized evaluation of prediction accuracy	Open-source Python package
Precursor Libraries	Commercial precursor databases, text-mined template sets [20]	Ground truth for precursor prediction models	Domain-specific curation

Implications and Future Directions

The development of CSLLM and similar frameworks has profound implications for accelerating functional materials discovery. By bridging the gap between computational design and experimental synthesis, these approaches can significantly increase the success rate of materials discovery campaigns.

Future research directions likely include:

• Integration of synthesis condition prediction (temperature, time, atmosphere)
• Extension to broader material classes including metastable and disordered phases
• Incorporation of active learning for continuous model improvement from experimental results
• Development of multi-modal approaches combining textual, graph-based, and structural representations

As LLMs continue to evolve and materials datasets expand, the accuracy and scope of synthesis prediction frameworks will undoubtedly improve, potentially transforming materials discovery from an empirical art to a predictive science.

The CSLLM framework represents a transformative approach to the long-standing challenge of predicting crystal synthesizability. By leveraging large language models fine-tuned on comprehensive materials data, CSLLM achieves unprecedented accuracy in synthesizability assessment while also providing practical guidance on synthesis methods and precursors. This capability addresses a critical bottleneck in computational materials discovery, potentially accelerating the translation of theoretical predictions to laboratory realization. As the field advances, integration of such predictive frameworks into materials design workflows promises to significantly enhance the efficiency and success rate of materials discovery for applications ranging from energy storage to pharmaceutical development.

The discovery of novel inorganic crystalline materials is fundamentally bottlenecked by the challenge of predicting synthesizable compounds and their viable synthesis pathways. While traditional methods rely on chemical intuition and expensive trial-and-error, artificial intelligence presents a transformative opportunity. This technical guide explores the integration of Element-Wise Graph Neural Networks—inspired by Kolmogorov-Arnold Networks (KANs)—as a powerful framework for retrosynthetic analysis of inorganic crystals. By systematically embedding learnable univariate functions across node embedding, message passing, and readout components, KA-GNNs achieve unprecedented performance in predicting stable, synthesizable materials, as demonstrated by the discovery of over 381,000 new stable crystals in recent large-scale implementations. We provide comprehensive methodological protocols, quantitative benchmarking, and implementation tools to equip researchers with cutting-edge capabilities for accelerating materials discovery.

The targeted synthesis of crystalline inorganic materials represents a grand challenge in materials science and solid-state chemistry, complicated by the absence of well-understood reaction mechanisms that typically guide organic synthesis [3]. Unlike organic molecules that can often be synthesized through sequence-based reactions, inorganic materials require consideration of thermodynamic stabilization, kinetic pathways, and complex solid-state interactions [3]. Despite computational advances, reliably identifying synthesizable inorganic crystalline materials remains an unsolved problem critical for realizing autonomous materials discovery.

Current approaches for predicting synthesizability face several fundamental limitations:

• Charge-Balancing Inadequacy: The commonly employed charge-balancing criterion, which filters materials based on net neutral ionic charge, fails to accurately predict synthesizability, correctly identifying only 37% of known synthesized inorganic materials and a mere 23% of known ionic binary cesium compounds [3].

• Thermodynamic Stability Limitations: Density functional theory (DFT) calculations of formation energy and decomposition enthalpy capture only approximately 50% of synthesized inorganic crystalline materials, failing to account for kinetic stabilization and non-thermodynamic factors [3].

• Data Scarcity: Experimental melting point data, a crucial property for synthesis planning, remains scarce due to measurement challenges, with only 799 well-characterized inorganic crystals available in standard references [33].

• Human Expertise Bottlenecks: Expert solid-state chemists specializing in specific synthetic techniques require extensive time for evaluation, creating significant throughput limitations in materials exploration [3].

The development of graph neural networks for materials discovery has begun to overcome these challenges through large-scale active learning frameworks. Notably, the GNoME (graph networks for materials exploration) project has demonstrated the capability to discover 2.2 million potentially stable structures, expanding the number of known stable crystals by almost an order of magnitude [34]. However, the critical task of predicting viable synthesis pathways for these discovered materials requires more sophisticated approaches that can capture the complex relationships between elemental composition, crystal structure, and synthetic accessibility.

Element-Wise Graph Neural Networks: Theoretical Framework and Architecture

Element-Wise Graph Neural Networks represent a significant architectural innovation through the integration of Kolmogorov-Arnold Network (KAN) principles into geometric deep learning frameworks. Inspired by the Kolmogorov-Arnold representation theorem, KANs replace traditional multilayer perceptrons (MLPs) with learnable univariate functions positioned on edges rather than nodes, enabling more accurate and interpretable modeling of complex scientific relationships [35].

Core Mathematical Foundation

The Kolmogorov-Arnold superposition theorem states that any multivariate continuous function can be expressed as a finite composition of univariate functions and additions:

$$f(\mathbf{x}) = f(x1, \ldots, xn) = \sum{q=1}^{2n+1} \Phiq \left( \sum{p=1}^n \phi{q,p}(x_p) \right)$$

where $\phi{q,p}$ and $\Phiq$ represent univariate functions. This theoretical foundation enables KANs to achieve more compact and accurate function approximations with smoother gradients compared to traditional MLPs [35].

In the context of GNNs for materials science, Fourier-based univariate functions have demonstrated particular effectiveness for capturing both low-frequency and high-frequency structural patterns in crystal graphs. The Fourier-KAN formulation implements these pre-activation functions as:

$$\text{KAN}(x) = \sum{k=1}^N \left( ak \sin(kx) + b_k \cos(kx) \right)$$

This Fourier-based approach provides strong theoretical approximation guarantees grounded in Carleson's convergence theorem and Fefferman's multivariate extension, enabling the model to accurately represent complex multivariate functions relevant to materials property prediction [35].

KA-GNN Architectural Integration

The KA-GNN framework systematically integrates KAN modules across three fundamental components of graph neural networks:

• Node Embedding: Atomic features (atomic number, radius) and neighboring bond features (bond type, length) are concatenated and processed through a KAN layer to generate initial node embeddings that encode both atomic identity and local chemical context [35].

• Message Passing: Traditional aggregation functions are replaced with KAN-based transformations that modulate feature interactions during message passing, enhancing the expressiveness of neighborhood information propagation [35].

• Graph-Level Readout: KAN modules generate more expressive graph-level representations by capturing complex, non-linear relationships in the aggregated features, replacing conventional MLP-based readout functions [35].

Table 1: KA-GNN Architectural Components and Their Functions

Component	Traditional Approach	KA-GNN Implementation	Key Advantage
Node Embedding	MLP with fixed activations	Fourier-KAN layer with atomic and bond features	Data-dependent trigonometric transformations
Message Passing	Weighted sum aggregation	KAN-modulated feature interactions	Enhanced expressiveness in neighborhood aggregation
Readout Function	Global pooling + MLP	KAN-based composition	Captures complex non-linear relationships
Residual Connections	Linear transformations	Residual KAN blocks	Improved gradient flow and training dynamics

Quantitative Benchmarking and Performance Analysis

KA-GNN frameworks have demonstrated remarkable empirical performance across multiple materials discovery benchmarks, consistently outperforming conventional GNN architectures in both prediction accuracy and computational efficiency.

Performance Metrics and Comparative Analysis

Recent large-scale evaluations across seven molecular benchmarks show that KA-GNNs achieve significant improvements over conventional GNNs [35]. Through active learning cycles, these models have improved from initial hit rates of less than 6% (structural) and 3% (compositional) to final precision exceeding 80% for structure-based predictions and 33% per 100 trials for composition-only predictions [34].

Table 2: Performance Comparison of Materials Prediction Models

Model/Approach	Prediction Task	Key Metric	Performance	Reference
KA-GNN (Fourier)	Molecular property prediction	Accuracy	Consistent outperformance vs. conventional GNNs	[35]
GNoME (Active Learning)	Crystal stability prediction	Hit rate	>80% (structure), 33% (composition)	[34]
SynthNN	Synthesizability classification	Precision	7× higher than DFT formation energy	[3]
Charge-Balancing	Synthesizability prediction	Accuracy	37% on known materials	[3]
DFT Formation Energy	Stability prediction	Coverage	50% of synthesized materials	[3]
GeoCGNN (Transfer Learning)	Melting point prediction	RMSE	218 K (46% improvement)	[33]

The GNoME framework, which utilizes scaled GNNs, exemplifies the power of these approaches, having discovered 381,000 new stable crystals on the updated convex hull—an order-of-magnitude expansion from previous knowledge [34]. These models exhibit emergent out-of-distribution generalization, accurately predicting structures with 5+ unique elements despite their omission from training data [34].

Synthesizability Prediction Breakthroughs

For the specific task of predicting synthesizability—a more challenging objective than stability prediction—specialized models like SynthNN have demonstrated remarkable capabilities. In head-to-head material discovery comparisons against 20 expert material scientists, SynthNN outperformed all human experts, achieving 1.5× higher precision and completing the task five orders of magnitude faster than the best-performing human expert [3].

Notably, without any prior chemical knowledge, SynthNN learns fundamental chemical principles including charge-balancing, chemical family relationships, and ionicity directly from the distribution of synthesized materials in the Inorganic Crystal Structure Database (ICSD) [3]. This demonstrates the powerful knowledge extraction capabilities of properly architected deep learning models for materials science.

Experimental Protocols and Methodologies

KA-GNN Implementation Workflow

Implementing effective Element-Wise Graph Neural Networks for retrosynthetic analysis requires careful attention to architectural details and training methodologies.

Candidate Generation: Two primary frameworks exist for generating candidate materials:

• Structural Candidates: Generated through symmetry-aware partial substitutions (SAPS) of available crystals, creating diverse structural variations while preserving crystal symmetry [34].
• Compositional Candidates: Generated through oxidation-state balancing with relaxed constraints, followed by structure initialization using ab initio random structure searching (AIRSS) [34].

KA-GNN Processing: The core architecture involves:

• Graph Representation: Crystals are represented as graphs with atoms as nodes and bonds as edges, with features including atomic number, radius, orbital configuration, bond type, and bond length [35] [34].
• Fourier-KAN Layers: Implemented using sine and cosine basis functions with tunable harmonic coefficients, typically bounded by a maximum harmonic parameter K for computational efficiency [35].
• Uncertainty Quantification: Implemented through deep ensembles to assess prediction confidence and guide active learning [34].

DFT Verification: Predicted stable candidates are verified using density functional theory calculations, typically performed with the Vienna Ab initio Simulation Package (VASP) using standardized settings from the Materials Project [34].

Transfer Learning for Enhanced Prediction

For properties with limited experimental data, such as melting temperature, transfer learning has proven highly effective. The protocol involves:

• Pre-training: Models are first trained on large computational databases of related properties, such as atomization energies of approximately 36,000 materials from the Materials Project [33].
• Fine-tuning: The pre-trained models are subsequently fine-tuned using smaller experimental datasets (e.g., 799 crystals with measured melting temperatures) [33].

This approach has demonstrated 46% improvement in prediction accuracy for melting temperatures compared to training from scratch, decreasing RMSE from 407 K to 218 K [33]. The effectiveness depends on the physical relationship between pre-training and target properties, with atomization energy showing stronger correlation with melting temperature than formation energy or band gap energy [33].

Table 3: Essential Resources for KA-GNN Implementation

Resource Category	Specific Tools/Databases	Function/Purpose	Access Method
Materials Databases	Materials Project (MP), Inorganic Crystal Structure Database (ICSD), Open Quantum Materials Database (OQMD)	Source of known crystal structures and properties for training and benchmarking	Public web portals and APIs
DFT Computational Tools	Vienna Ab initio Simulation Package (VASP)	Verification of predicted stable materials through energy calculations	Academic licensing
Candidate Generation	Symmetry-Aware Partial Substitutions (SAPS), Ab Initio Random Structure Searching (AIRSS)	Generation of diverse candidate structures for evaluation	Custom implementation
GNN Frameworks	PyTor Geometric, Deep Graph Library	Implementation of graph neural network architectures	Open source Python libraries
KA-GNN Specialized Components	Fourier-KAN layers, Message passing with learnable univariate functions	Enhanced expressivity and interpretability in graph networks	Custom implementation based on [35]
Active Learning Infrastructure	Deep ensembles, Uncertainty quantification, Automated DFT workflows	Iterative model improvement through targeted data acquisition	Custom pipeline implementation

Element-Wise Graph Neural Networks represent a paradigm shift in retrosynthetic analysis for inorganic materials, addressing fundamental challenges in synthesizability prediction through innovative architectural principles. By integrating Kolmogorov-Arnold Networks with geometric deep learning, KA-GNNs achieve unprecedented accuracy in identifying stable, synthesizable materials while providing enhanced interpretability through their learnable univariate functions.

The demonstrated capabilities of these models—from discovering hundreds of thousands of new stable crystals to outperforming human experts in synthesizability prediction—highlight their transformative potential for accelerating materials discovery. As these approaches continue to evolve through scaling laws, improved active learning strategies, and more sophisticated transfer learning techniques, they promise to fundamentally reshape how we explore and synthesize novel inorganic materials for technological applications across energy storage, electronics, and beyond.

Future research directions should focus on integrating kinetic synthesis factors, incorporating real-time experimental feedback, and developing more sophisticated multi-property optimization frameworks to further bridge the gap between computational prediction and experimental realization.

The discovery of novel inorganic crystalline materials is a fundamental driver of technological advancement. While computational methods, particularly density functional theory (DFT), have successfully identified millions of candidate materials with promising properties, a significant bottleneck remains: bridging the gap between theoretical prediction and experimental realization [7]. The central challenge lies in accurately predicting crystallographic synthesizability—whether a hypothetical crystal structure can be experimentally synthesized—and determining the practical synthesis pathways to achieve it.

Traditional approaches for assessing synthesizability rely on thermodynamic or kinetic stability metrics, such as formation energy and energy above the convex hull [7]. However, these methods often prove insufficient; numerous structures with favorable formation energies remain unsynthesized, while various metastable structures are routinely synthesized despite less favorable thermodynamics [7]. This discrepancy highlights that synthesizability is influenced by a complex array of factors beyond thermodynamic stability, including precursor selection, reaction conditions, and kinetic barriers [3]. This whitepaper examines the integration of synthesizability prediction with precursor and method recommendation, framing it within the broader challenge of predicting synthesizable inorganic crystals.

Core Methodologies and Quantitative Benchmarks

Data Curation and Representation for Machine Learning

A critical first step involves constructing comprehensive datasets of both synthesizable and non-synthesizable materials for model training.

• Positive Data Sources: The Inorganic Crystal Structure Database (ICSD) serves as a reliable source of experimentally validated, synthesizable crystal structures [7] [3]. One methodology involves selecting structures with up to 40 atoms and seven different elements, excluding disordered phases to focus on ordered crystals [7].
• Negative Data Construction: Creating a robust set of non-synthesizable examples is a known challenge. One advanced approach uses a pre-trained Positive-Unlabeled (PU) learning model to assign a CLscore to theoretical structures from databases like the Materials Project (MP) and OQMD. Structures with the lowest CLscores (e.g., <0.1) are selected as high-confidence negative examples, effectively building a balanced dataset [7].
• Crystal Structure Representation: For efficient processing by language models, crystal structures must be converted into a compact text format. The material string representation condenses essential crystallographic information—space group, lattice parameters, and atomic coordinates with Wyckoff positions—into a human-readable format, eliminating redundancies found in CIF or POSCAR files [7].

Predictive Modeling Architectures

Recent advances have produced specialized models that address different aspects of the synthesis prediction problem. The table below summarizes the performance of key models.

Table 1: Performance Benchmarks of Key Synthesizability Prediction Models

Model Name	Input Type	Primary Task	Reported Accuracy	Key Advantage
CSLLM (Synthesizability LLM) [7]	Crystal Structure (Text)	Synthesizability Classification	98.6%	High accuracy & generalization on complex structures
SynthNN [3]	Chemical Composition	Synthesizability Classification	7x higher precision than DFT formation energy	Operates without structural information
CSLLM (Method LLM) [7]	Crystal Structure (Text)	Synthetic Method Classification	91.0%	Recommends solid-state or solution routes
CSLLM (Precursor LLM) [7]	Crystal Structure (Text)	Precursor Identification	80.2% success	Identifies solid-state precursors for binary/ternary compounds
Rank-Average Ensemble [22]	Composition & Structure	Synthesizability Scoring	Successful experimental synthesis of 7/16 targets	Combines compositional and structural signals for enhanced ranking

An Integrated Framework: From Crystal Structure to Synthesis Recipe

The most robust systems integrate multiple specialized models into a cohesive pipeline. The Crystal Synthesis Large Language Models (CSLLM) framework exemplifies this approach, employing three distinct LLMs fine-tuned for synthesizability prediction, method recommendation, and precursor identification [7]. The following diagram illustrates the integrated workflow from a candidate structure to a proposed synthesis recipe.

Workflow Execution

• Synthesizability Assessment: The candidate crystal structure, represented as a material string, is processed by the Synthesizability LLM. This model classifies the structure as synthesizable or non-synthesizable with high accuracy (98.6%), significantly outperforming traditional stability metrics [7].
• Synthetic Method Recommendation: For structures deemed synthesizable, the Method LLM classifies the most probable synthetic pathway, such as solid-state or solution methods, achieving 91.0% accuracy [7]. This step is crucial as it dictates the experimental setup and conditions.
• Precursor Identification: Finally, the Precursor LLM identifies suitable solid-state precursor compounds for the target material. This model achieves an 80.2% success rate for common binary and ternary compounds [7]. Advanced pipelines can further refine this step by employing models like Retro-Rank-In to rank viable precursors and SyntMTE to predict critical calcination temperatures [22].

Experimental Validation and Synthesis Protocol

The ultimate validation of any predictive pipeline is its success in guiding the experimental synthesis of novel materials. One study screened over 4.4 million computational structures using a combined synthesizability score, identifying approximately 500 high-priority candidates [22]. After precursor planning and filtering, this led to experimental attempts on 16 targets.

Table 2: Experimental Reagents and Materials for Solid-State Synthesis

Research Reagent / Material	Function in Synthesis	Experimental Considerations
Solid-State Precursors	Provide the elemental components for the target material.	Purity, particle size, and reactivity are critical. Selected via precursor-suggestion models [22].
High-Temperature Furnace	Provides the thermal energy required for solid-state reaction and diffusion.	Must achieve and maintain precise temperatures (e.g., predicted calcination temperature) [22].
X-Ray Diffractometer (XRD)	Characterizes the crystalline structure of the synthesis product.	Used for verification by comparing experimental and target diffraction patterns [22].
Ball Mill	Homogenizes precursor powders to increase reactivity.	Ensures intimate mixing of precursors for a complete reaction.

The experimental protocol for a solid-state synthesis, as derived from the validated pipeline, is as follows [22]:

• Precursor Selection and Weighing: Input the target compound's formula into the precursor recommendation model (e.g., Precursor LLM or Retro-Rank-In) to receive a ranked list of viable solid-state precursors. Weigh out the precursors according to the stoichiometric ratios of the balanced reaction.
• Mechanical Milling: Transfer the precursor powders into a ball mill jar. Mill for 15-60 minutes to ensure thorough mixing and a reduction in particle size, which increases the surface area for reaction.
• Calcination (Heating): Place the homogenized powder into a high-temperature crucible (e.g., alumina). Insert the crucible into a furnace. Heat to the model-predicted calcination temperature (e.g., using SyntMTE) at a controlled ramp rate (e.g., 5°C/min). Hold at the target temperature for a duration of 6-24 hours to allow for complete reaction and crystal formation.
• Product Characterization: After the sample cools to room temperature, characterize the resulting powder using X-ray diffraction (XRD). Compare the measured diffraction pattern to the pattern simulated from the target crystal structure to verify successful synthesis.

This protocol resulted in the successful synthesis and characterization of 7 out of 16 target materials, including one completely novel compound, demonstrating the practical efficacy of the integrated pipeline [22].

The integration of synthesizability prediction with precursor and method recommendation represents a paradigm shift in computational materials discovery. By moving beyond thermodynamic stability to model the complex, multi-factor nature of chemical synthesis, frameworks like CSLLM provide an actionable bridge from theoretical simulation to experimental realization. The successful experimental validation of these computational pipelines confirms their potential to dramatically accelerate the discovery and development of new functional inorganic materials. Future progress will depend on expanding and refining the datasets for synthesis routes and further improving the explainability of model predictions to build greater trust and utility for experimental chemists.

Overcoming Data Scarcity, Model Explainability, and Hallucination

The accurate prediction of which hypothetical inorganic crystals can be successfully synthesized represents a fundamental challenge in materials science and drug development. While computational models can generate millions of candidate structures with desirable properties, the vast majority may not be synthetically accessible, making experimental validation impractical. The development of reliable machine learning (ML) models for synthesizability prediction depends critically on the quality and composition of the training datasets used. Constructing representative datasets containing both confirmed synthesizable crystals and validated non-synthesizable examples presents unique methodological challenges that directly impact model performance and real-world applicability.

This technical guide examines current approaches for curating positive and non-synthesizable crystal structure datasets within the broader context of predicting synthesizable inorganic crystals. We detail specific protocols for data collection, labeling, and representation, providing researchers with practical methodologies for dataset construction. By addressing the fundamental data challenges in this field, we aim to establish robust foundations for the next generation of synthesizability prediction models that can more effectively bridge the gap between computational materials design and experimental realization.

Data Curation Methodologies and Experimental Protocols

Sourcing Positive (Synthesizable) Crystal Structures

Well-established experimental crystallographic databases serve as the primary sources for confirmed synthesizable crystal structures. These databases provide chemically diverse, experimentally verified structures that can be used as positive examples in training datasets.

Table 1: Primary Data Sources for Positive Examples

Database	Content Focus	Data Volume	Key Considerations
Inorganic Crystal Structure Database (ICSD)	Experimentally synthesized inorganic crystals	~70,000 structures after filtering [17]	Exclude disordered structures; apply composition/size filters (e.g., ≤40 atoms, ≤7 elements)
Crystallography Open Database (COD)	Open-access collection of crystal structures	3000+ samples in curated sets [36]	Include distinct polymorphs for chemical compositions also represented in negative class
Materials Project (MP)	Computationally characterized materials	38,347 synthesized structures in processed sets [2]	Structures are typically derived from ICSD; apply text description length filters for LLM applications

Standardized filtering protocols must be applied to ensure data quality and manage computational complexity. For inorganic crystals, common filters include: removing structures with disorder; limiting structures to those containing no more than 40 atoms and seven different elements [17]; and excluding structures where text descriptions exceed character limits for natural language processing applications [2]. Including all distinct polymorphs for chemical compositions that also appear in the negative class significantly enhances model performance by providing necessary structural information for learning distinctions between classes [36].

Constructing Negative (Non-Synthesizable) Crystal Classes

A fundamental challenge in synthesizability prediction is the lack of definitively non-synthesizable examples, as unsuccessful syntheses are rarely reported. Researchers have developed several methodological approaches to construct representative negative classes, each with distinct advantages and limitations.

Table 2: Methodologies for Negative Class Construction

Method	Underlying Principle	Dataset Scale	Performance
Positive-Unlabeled (PU) Learning	Treats hypothetical structures as unlabeled; uses CLscore threshold (<0.1) to identify non-synthesizable candidates [17]	80,000 from 1.4M theoretical structures [17]	98.6% accuracy in synthesizability prediction [17]
Crystal Anomaly Selection	Selects unobserved structures for well-studied compositions (>3,306 literature mentions) [36]	600 anomalies from 108 compositions [36]	Enables binary classification across diverse crystal types
Charge-Balancing Filter	Filters out materials without net neutral ionic charge using common oxidation states [3]	N/A	Limited accuracy (37% of synthesized materials are charge-balanced) [3]

The Positive-Unlabeled (PU) learning approach has demonstrated particularly strong performance. This method calculates a CLscore for hypothetical structures from sources like the Materials Project, with scores below 0.1 indicating high probability of non-synthesizability [17]. To create balanced datasets, researchers select structures with the lowest CLscores, with verification that 98.3% of positive examples have CLscores greater than 0.1, validating the threshold choice [17].

The crystal anomaly approach identifies frequently studied chemical compositions (top 0.1% with ≥3306 literature mentions) and designates their unobserved crystal structures as anomalies, based on the assumption that extensively studied compositions have likely had all synthesizable structures discovered [36]. The number of generated anomaly structures is typically balanced with synthesizable structures for each composition, with at least five unobserved structures generated per composition.

Data Representation Formats for Machine Learning

Effective representation of crystal structures is essential for ML model training. Different representation formats enable various computational approaches to synthesizability prediction.

• CIF and POSCAR Formats: Traditional structural representations containing lattice parameters, atomic coordinates, and symmetry information. These can be processed by graph neural networks or converted to other representations [17].

• Text Descriptions: Human-readable crystal structure descriptions generated by tools like Robocrystallographer enable fine-tuning of large language models (LLMs). These descriptions are particularly effective when combined with text-embedding models for feature extraction [2].

• 3D Voxel Images: Color-coded three-dimensional images representing atomic structures and chemical attributes, suitable for convolutional neural networks. This representation simultaneously captures structural and chemical features across diverse crystal types [36].

• Material Strings: Efficient text representations that integrate essential crystal information while eliminating redundancy from full structural listings. These specialized formats optimize LLM fine-tuning by focusing on salient features [17].

Integrated Workflows for Dataset Construction and Model Training

Combining data curation methodologies with appropriate representations enables complete workflows for synthesizability prediction. The following diagram illustrates a comprehensive pipeline from raw data to predictions and explanations.

This integrated workflow demonstrates how curated datasets enable multiple prediction capabilities. The Crystal Synthesis Large Language Models (CSLLM) framework exemplifies this approach, utilizing three specialized LLMs to predict synthesizability, identify appropriate synthetic methods (solid-state or solution), and suggest suitable precursors [17]. This comprehensive system achieves 98.6% accuracy in synthesizability prediction and exceeds 90% accuracy in method classification and precursor identification for common compounds [17].

Table 3: Computational Tools for Dataset Construction and Synthesizability Prediction

Tool/Resource	Function	Application Context
DASH	Crystal structure solution from powder diffraction data	Experimental structure determination [37]
TOPAS-Academic	Rietveld refinement of powder diffraction data	Experimental structure validation [37]
Robocrystallographer	Text description generation from crystal structures	LLM input preparation [2]
PU Learning Algorithms	Identification of non-synthesizable structures from hypothetical databases	Negative class construction [17] [3]
Convolutional Auto-encoders	Feature learning from 3D crystal images	Unsupervised representation learning [36]
Fine-tuned LLMs (GPT-4o-mini)	Synthesizability classification from text structure descriptions	Explainable synthesizability prediction [2]
Universal Interatomic Potentials	Rapid energy estimation for crystal stability screening	Pre-filtering for thermodynamic stability [9]

The construction of representative datasets for crystal synthesizability prediction remains both challenging and essential for advancing materials discovery. Methodologies for curating negative classes, particularly through PU learning and crystal anomaly selection, have demonstrated significant improvements in prediction accuracy over traditional thermodynamic approaches. The integration of diverse data representation formats—from graph networks to text descriptions—enables researchers to leverage increasingly sophisticated ML architectures.

Future progress will likely depend on addressing several persistent challenges: expanding the scope of definitively non-synthesizable examples, improving cross-domain generalization, and enhancing the explainability of model predictions. Standardized benchmarking frameworks like Matbench Discovery will be crucial for objectively evaluating new approaches across diverse chemical spaces [9]. As these methodologies mature, robust dataset construction practices will play an increasingly critical role in bridging the gap between computational materials design and experimental synthesis, ultimately accelerating the discovery of novel functional materials for pharmaceutical and technological applications.

The discovery of new functional materials is a cornerstone of technological advancement, driving innovation in fields from renewable energy to healthcare. Computational methods, particularly density functional theory (DFT) and machine learning, have dramatically accelerated the identification of candidate materials with promising properties. However, a significant bottleneck remains: the experimental validation of these hypothetical compounds. A critical unsolved challenge in computational materials science is reliably predicting which theoretically designed crystals are synthetically accessible—a property known as synthesizability [3].

The core obstacle in developing data-driven synthesizability predictors is the fundamental nature of available materials data. While databases like the Inorganic Crystal Structure Database (ICSD) contain thousands of experimentally realized structures (positive examples), they contain virtually no confirmed negative examples (unsynthesizable materials) [3] [38]. Failed synthesis attempts are rarely published, creating a severe data asymmetry [14] [38]. Traditional supervised learning requires both positive and negative examples to train a classifier, making this paradigm unsuitable for synthesizability prediction.

Positive-Unlabeled (PU) learning has emerged as a powerful semi-supervised framework to address this exact challenge, enabling model training where only positive and unlabeled examples are available [14]. This paradigm is particularly valuable in materials informatics, where it allows researchers to learn from the distribution of known synthesized materials while making inferences about the vast space of hypothetical, unlabeled compounds.

Core Principles of PU Learning

Problem Formulation and Key Assumptions

In the context of synthesizability prediction, PU learning treats all experimentally verified crystals from databases like ICSD as positive (P) examples. The unlabeled (U) set consists of hypothetical crystals from computational databases like the Materials Project that lack experimental validation; this set contains both synthesizable and unsynthesizable materials, but their labels are unknown [3] [13].

PU learning algorithms rely on two key assumptions:

• Selected Completely at Random (SCAR) Assumption: The probability that a positive example is labeled (i.e., included in the positive set) is independent of its features [38].
• Positive examples are distinguishable from negative examples based on their feature distributions, allowing a classifier to separate them given sufficient data.

Common PU Learning Strategies

Several strategic approaches have been developed for PU learning:

• Two-Step Techniques: These methods first identify reliable negative examples from the unlabeled set, then iteratively refine a classifier using these identified negatives along with the positive examples [38].
• Biased Learning: This approach treats all unlabeled examples as negative but assigns lower weights to them during training to account for the contamination of the negative set with positive examples [3].
• Class Prior Incorporation: Some methods incorporate estimates of the proportion of positive examples in the unlabeled data to improve classification accuracy.

Quantitative Performance of PU Learning Models

Recent research has demonstrated the effectiveness of PU learning across various material classes and prediction tasks. The table below summarizes key performance metrics from recent studies.

Table 1: Performance Comparison of Recent PU Learning Models for Synthesizability Prediction

Model Name	Material Scope	Architecture	Key Performance Metrics	Reference
CSLLM	3D crystal structures	Specialized Large Language Model	98.6% accuracy on test set	[7]
CPUL	Virtual crystals	Contrastive Learning + PU Learning	93.95% true positive rate on MP data	[13]
SynCoTrain	Oxide crystals	Dual GCNN co-training (ALIGNN + SchNet)	High recall on internal and leave-out test sets	[38]
PU-CGCNN	Inorganic crystals	Crystal Graph Convolutional Neural Network	87.4% true positive rate on test data	[39] [2]
PU-GPT-embedding	Inorganic crystals	LLM embeddings + PU classifier	Outperforms PU-CGCNN in true positive rate	[2]

Table 2: Comparison of PU Learning Against Traditional Synthesizability Metrics

Prediction Method	Basis of Prediction	Reported Accuracy/Performance	Limitations
Energy Above Hull (E hull)	Thermodynamic stability	Captures only ~50% of synthesized materials	Fails to account for kinetic stabilization and synthesis conditions [14] [3]
Charge-Balancing	Net neutral ionic charge	Only 37% of known synthesized materials are charge-balanced	Inflexible for different bonding environments [3]
PU Learning Models	Data-driven patterns from known materials	Up to 98.6% accuracy (CSLLM) [7]	Requires careful handling of unlabeled set contamination

Experimental Protocols and Methodologies

Data Curation and Preprocessing

A critical first step in applying PU learning to synthesizability prediction is the careful curation of datasets:

• Positive Set Construction: Extract experimentally synthesized crystals from reliable databases such as ICSD. For example, one study used 70,120 crystal structures from ICSD as positive examples, applying filters to exclude disordered structures and limit complexity (e.g., ≤40 atoms, ≤7 different elements) [7].

• Unlabeled Set Construction: Source hypothetical structures from computational databases like the Materials Project, Computational Materials Database, or Open Quantum Materials Database. The same study gathered 1,401,562 such structures and applied a pre-trained PU model to select 80,000 with the lowest crystal-likeness scores (CLscore <0.1) as the unlabeled set [7].

• Feature Representation: Convert crystal structures into machine-readable formats. Common approaches include:

  • Crystal Graph Representations: Represent crystals as graphs with atoms as nodes and bonds as edges, incorporating atomic coordinates and bond distances [38].
  • Text-Based Representations: Use tools like Robocrystallographer to generate textual descriptions of crystal structures, which can be processed by large language models [2].
  • Material Strings: A condensed text format capturing space group, lattice parameters, atomic species, Wyckoff positions, and coordinates [7].

Model Implementation Frameworks

Different architectural frameworks have been employed for PU learning in materials science:

Graph Neural Network Approaches:

Diagram 1: GNN-based PU Learning Workflow

• Architecture: Implement Crystal Graph Convolutional Neural Networks (CGCNN) or similar architectures that operate directly on crystal graphs [40] [2].
• Training: Apply biased learning where unlabeled examples are treated as negative but with lower weights, or use iterative self-training to identify reliable negatives from the unlabeled set.
• Implementation: The model is typically trained to output a crystal-likeness score (CLscore) between 0 and 1, with higher scores indicating higher predicted synthesizability [13].

Large Language Model Approaches:

Diagram 2: LLM-based PU Learning Workflow

• Architecture: Utilize pre-trained LLMs like GPT-4, fine-tuned on text representations of crystal structures [7] [2].
• Training: Generate text embeddings from crystal structure descriptions, then train a binary PU classifier neural network on these embeddings.
• Implementation: The fine-tuned model can predict synthesizability and, importantly, generate human-readable explanations for its predictions [2].

Dual Classifier Co-Training Framework (SynCoTrain):

Diagram 3: Dual Classifier Co-Training

• Architecture: Employ two complementary graph neural networks (e.g., ALIGNN and SchNet) that learn from different structural perspectives [38].
• Training: Implement iterative co-training where classifiers exchange predictions on unlabeled data, gradually refining the decision boundary while mitigating individual model biases.
• Implementation: Final predictions are based on the average of both classifiers' outputs, improving robustness and generalization [38].

Model Validation Strategies

Validating PU learning models presents unique challenges due to the absence of true negative examples:

• Hold-Out Testing: Reserve a subset of positive examples as a test set to evaluate the true positive rate (recall) [13] [2].
• α-Estimation: Use statistical methods to approximate precision and false positive rates despite the lack of confirmed negatives [2].
• External Validation: Test model predictions against experimental results, as demonstrated in studies that synthesized predicted compounds [14].
• Cross-Validation: Implement k-fold cross-validation on the positive and unlabeled sets to assess model stability.

Table 3: Key Computational Tools and Databases for PU Learning in Materials Science

Resource Name	Type	Primary Function	Application in PU Learning
Materials Project (MP)	Database	Repository of computed materials properties	Source of unlabeled hypothetical structures [14] [13]
Inorganic Crystal Structure Database (ICSD)	Database	Curated experimental crystal structures	Source of confirmed positive examples [7] [3]
Pymatgen	Software Library	Materials analysis	Processing crystal structures and materials data [14]
Robocrystallographer	Software Tool	Text description of crystal structures	Generating LLM-readable structure representations [2]
CGCNN	Framework	Graph neural networks for crystals	Building base models for structure-based prediction [40]
ALIGNN	Framework	Graph neural networks incorporating angles	Enhanced structure representation for co-training [38]

Future Directions and Challenges

While PU learning has demonstrated remarkable success in synthesizability prediction, several challenges remain. Model generalizability across diverse material families needs improvement, particularly for compounds with complex bonding environments or those requiring specialized synthesis techniques [38] [41]. The development of explainable AI approaches integrated with PU learning will be crucial for building trust in predictions and providing chemical insights to guide experimental efforts [2].

Future research directions include hybrid approaches that combine physical knowledge with data-driven models, integration of synthesis condition prediction, and the creation of standardized benchmarks for evaluating synthesizability predictors [41]. As autonomous laboratories become more prevalent, PU learning models will play an increasingly important role in guiding experimental synthesis campaigns, ultimately accelerating the discovery of novel functional materials.

The prediction of synthesizable inorganic crystals represents a fundamental challenge in materials science, bridging the gap between computational design and experimental realization. While machine learning (ML) models like SynthNN have demonstrated remarkable capability in identifying synthesizable materials, their complex nature often renders them as "black boxes," limiting trust and practical adoption by researchers. This whitepaper explores the integration of Large Language Models (LLMs) as a powerful mechanism to generate natural language explanations for such predictive models. By translating complex model outputs—such as feature attributions from methods like SHAP—into accessible, human-readable narratives, LLMs enhance interpretability, foster trust, and facilitate actionable insights. Framed within the context of crystalline materials research, we provide a technical guide on methodologies, experimental protocols, and visualization tools to implement LLM-driven explainable AI (XAI), empowering scientists to better understand and utilize predictive synthesizability assessments.

The discovery of new inorganic crystalline materials is pivotal for technological advancement, yet a significant bottleneck lies in reliably predicting which computationally designed materials are synthetically accessible. Traditional approaches to assessing synthesizability have relied on expert intuition, charge-balancing criteria, or density functional theory (DFT)-calculated formation energies. However, these methods often fall short; for instance, charge-balancing fails to accurately predict a large portion of known synthesized materials, with only 37% of synthesized inorganic materials being charge-balanceable according to common oxidation states [3]. Furthermore, thermodynamic stability alone is an insufficient metric, as it fails to account for kinetic stabilization, experimental conditions, and complex chemical relationships [3] [16].

Machine learning models, such as deep learning synthesizability models (SynthNN), have emerged as powerful tools to overcome these limitations. Trained on extensive databases of known materials, these models learn the underlying chemical principles of synthesizability directly from data, outperforming both traditional computational methods and human experts. SynthNN, for example, has been shown to identify synthesizable materials with 7x higher precision than formation energy-based approaches and 1.5x higher precision than the best human expert, while operating five orders of magnitude faster [3] [23]. Despite their performance, the internal decision-making processes of these complex models remain opaque, creating a critical barrier to their reliable application in high-stakes research and development. This is where the fusion of XAI and LLMs becomes essential.

The Role of Explainable AI (XAI) in Materials Science

Explainable AI (XAI) encompasses techniques designed to make the outputs of AI systems understandable to humans, providing insight into the model's internal logic and the factors influencing its predictions [42] [43]. In materials science, this translates to understanding why a model classifies a specific chemical composition as synthesizable or not.

Common XAI techniques used in ML include:

• SHAP (SHapley Additive exPlanations): A game-theoretic approach that assigns each input feature (e.g., elemental properties in a composition) an importance value for a particular prediction [44] [43].
• LIME (Local Interpretable Model-agnostic Explanations): Explains individual predictions by approximating the complex model locally with a simpler, interpretable model [44] [43].
• Counterfactual Explanations: Identify the minimal changes required in the input to alter the model's output (e.g., "If element Y were replaced with Z, the material would be predicted as synthesizable") [44].

For ML models predicting material synthesizability, SHAP values might reveal that the model's decision was heavily influenced by the electronegativity difference and atomic radius of the constituent elements. However, presenting these results as raw feature importance scores or complex visualizations can be difficult for domain experts without ML expertise to interpret and act upon [45].

Large Language Models as Explanation Engines

Large Language Models, with their advanced natural language generation capabilities, offer a transformative solution to the interpretability challenge. They can be leveraged to automate the transformation of technical XAI outputs into coherent, natural language narratives, making insights accessible to a broader audience of researchers and stakeholders [42] [45].

The application of LLMs in XAI typically follows one of two paradigms:

• Post-hoc Explanation Generation: The LLM is provided with the model's prediction (e.g., "synthesizable") and the corresponding XAI output (e.g., a SHAP feature importance plot). It then synthesizes this information into a fluent, textual explanation [46].
• Integrated Reasoning: Techniques like Chain-of-Thought (CoT) prompting instruct the LLM to "think step-by-step," generating both the answer and the reasoning process in natural language. This provides a human-readable justification that can be debugged and evaluated [44].

A key system architecture, exemplified by MIT's EXPLINGO, divides this process into two components:

• NARRATOR: An LLM that converts structured explanation data (like SHAP values) into a paragraph of human-readable text, customizable by showing it a few example explanations to mimic a desired style [46].
• GRADER: A second LLM-based component that automatically evaluates the generated narrative on metrics like conciseness, accuracy, completeness, and fluency, providing a quality check for the end-user [46].

This approach limits the LLM's role to natural language generation, reducing the risk of introducing factual inaccuracies into the explanation while leveraging its fluency and adaptability [46].

Experimental Protocols and Data for Synthesizability Prediction

The development of robust synthesizability prediction models requires carefully curated data and structured experimental protocols. Below is a detailed methodology based on established approaches in the literature [3] [16].

Data Sourcing and Curation

• Primary Data Source: The Inorganic Crystal Structure Database (ICSD) is the foundational source for synthesizable (positive) examples, representing a nearly complete history of synthesized and structurally characterized crystalline inorganic materials [3] [16].
• Handling Unsynthesized Materials: Definitively labeling a material as unsynthesizable is problematic. The standard workaround is to generate a set of artificial, hypothetical chemical compositions that are not present in the ICSD. This creates a Positive-Unlabeled (PU) learning scenario, where the model learns from confirmed positive examples and a set of unlabeled examples that are presumed to be mostly negative [3].
• Feature Representation:
  • atom2vec: A learned atom embedding matrix optimized alongside the neural network parameters, which learns an optimal representation of chemical formulas directly from the distribution of synthesized materials without prior chemical knowledge [3].
  • Fourier-Transformed Crystal Properties (FTCP): A crystal representation that includes features in both real space and reciprocal space, capturing elemental properties and crystal periodicity [16].

Model Training and Validation

The following protocol outlines the training of a synthesizability classifier like SynthNN or an SC (Synthesizability Score) model:

• Data Partitioning: Split the data (ICSD positives + generated negatives) into training, validation, and test sets. A publication-year-split is particularly telling; the model is trained on data available until a certain year (e.g., 2015) and tested on materials reported in subsequent years (e.g., post-2019) to evaluate its predictive power for novel discoveries [16].
• Model Architecture: Implement a deep learning classifier, such as a convolutional neural network (CNN) designed to process the chosen feature representation (e.g., atom2vec, FTCP) [3] [16].
• PU-Learning Adjustment: During training, apply class-weighting to the unlabeled (artificially generated) examples according to their estimated likelihood of being synthesizable, to account for the noise in the negative class [3].
• Validation Metrics: Evaluate model performance using precision, recall, and F1-score. It is critical to acknowledge that the "false positives" in the test set may include synthesizable materials that simply have not been synthesized or recorded yet, which can depress precision metrics [3].

Table 1: Performance Comparison of Synthesizability Prediction Methods

Method	Key Principle	Precision	Recall	Key Advantage
Charge-Balancing [3]	Net neutral ionic charge	Very Low	Low	Chemically intuitive, fast
DFT Formation Energy [3]	Thermodynamic stability	Low	~50%	Physics-based
SynthNN (PU Learning) [3]	Data-driven classification	7x higher than DFT	High	Learns complex chemical relationships
SC Model (FTCP) [16]	Data-driven classification	82.6% (Ternary)	80.6% (Ternary)	Incorporates crystal structure info

Visualization of the LLM-XAI Workflow

The following diagram illustrates the integrated workflow of using an ML model for synthesizability prediction and an LLM to generate human-readable explanations.

Successful implementation of an LLM-driven explainability pipeline for materials prediction relies on a suite of computational and data resources.

Table 2: Essential Resources for LLM-XAI in Materials Research

Resource Name	Type	Function in the Pipeline
Inorganic Crystal Structure Database (ICSD) [3] [16]	Data Repository	Provides the foundational dataset of known, synthesized crystalline materials used for training and benchmarking.
Materials Project (MP) Database [16]	Data Repository	Supplies computationally derived material properties and structures, often used in conjunction with ICSD data.
SHAP/LIME [44] [43]	Explainable AI Library	Generates the primary, technical explanations (feature attributions) for the ML model's predictions.
Pre-trained Large Language Model (e.g., via LM Studio) [45]	Software Tool	Serves as the core engine for generating natural language narratives from XAI outputs; offline deployment protects data privacy.
Python Materials Genomics (pymatgen) [16]	Software Library	Provides robust tools for analyzing materials data, manipulating crystal structures, and generating input features for ML models.
Atom2Vec / FTCP [3] [16]	Representation Method	Transforms raw chemical compositions and crystal structures into numerical representations suitable for machine learning.

The convergence of synthesizability prediction models and LLM-powered explainability marks a significant leap toward trustworthy and actionable AI in materials science. By translating the opaque logic of high-performing black-box models into clear, natural language narratives, we empower researchers to understand, critique, and ultimately trust AI-generated insights. This synergy not only accelerates the validation of new materials but also deepens our fundamental understanding of the chemical principles governing synthesis. As LLM and XAI techniques continue to mature, their integration will be crucial for realizing the full potential of autonomous materials discovery, ensuring that these powerful systems are not just predictors, but collaborative partners in scientific innovation.

The application of Large Language Models (LLMs) to materials science represents a paradigm shift in the acceleration of materials discovery. However, their impressive generative capabilities come with a significant risk: the production of factually inaccurate or unsupported information, a phenomenon known as hallucination [47] [48]. In the high-stakes domain of inorganic crystal prediction, where experimental validation is resource-intensive, hallucinations can lead to substantial wasted resources and misguided research directions. Hallucinations are formally defined as instances where model-generated content is fluent and syntactically correct but is factually inaccurate, ungrounded, or inconsistent with source material or established knowledge [47] [48]. The probabilistic nature of LLMs, which prioritizes statistically likely token sequences over epistemic truth, makes hallucination a fundamental mathematical inevitability rather than a simple correctable error [48]. Within materials science, this manifests in unique ways, such as proposing thermodynamically unstable crystal structures, non-existent synthesis pathways, or materials with contradictory physical properties [7] [49]. Combating these illusions is therefore a prerequisite for developing trustworthy AI partners in scientific discovery, forming a critical foundation for overcoming the fundamental challenges in predicting synthesizable inorganic crystals.

Defining and Categorizing Hallucinations in Materials Prediction

Understanding the specific forms of hallucination in materials science is essential for developing targeted mitigation strategies. The general taxonomy of LLM hallucinations can be adapted to the domain-specific challenges of crystal structure and synthesis prediction.

Table 1: Taxonomy of Hallucinations in LLM-Based Materials Prediction

Category	Subtype	Manifestation in Materials Science
Intrinsic (Factuality Errors)	Entity-Error Hallucinations	Generating non-existent crystal structures or inventing precursor chemicals with no CAS registry number [49] [48].
	Relation-Error Hallucinations	Proposing syntheses with incorrect temperature parameters or suggesting chronologically impossible discovery claims [48].
	Outdatedness Hallucinations	Recommending superseded synthetic methods or using outdated material property databases [48].
Extrinsic (Faithfulness Errors)	Unverifiability Hallucinations	Proposing precursor sets that violate charge neutrality or cannot be traced to reliable sources [49] [48].
	Incompleteness Hallucinations	Omitting critical synthesis conditions, such as atmospheric controls, from a generated recipe [48].

The following diagram illustrates the logical relationships between the major hallucination categories and their specific subtypes as they manifest in materials science applications.

Root Causes: Why Hallucinations Occur in Materials LLMs

The propensity for LLMs to hallucinate in materials science applications stems from vulnerabilities across the entire model development lifecycle. The primary causes can be categorized into data-related, training-related, and inference-related factors.

• Data Collection and Preparation Flaws: The quality of training data is a foundational factor. LLMs trained on large, unfiltered internet corpora ingest scientific information of varying reliability. In materials science, a significant challenge is the relative scarcity of structured data; the available data for crystals (10^5–10^6) is vastly smaller than for organic molecules (10^8–10^9) [7]. This data sparsity can force models to extrapolate beyond their knowledge. Furthermore, training data often suffers from underrepresentation of negative results and failed syntheses, creating a biased view of chemical reality [41].

• Training and Architectural Limitations: The next-token prediction objective, central to LLM pre-training, prioritizes linguistic plausibility over factual accuracy [47] [48]. A model may thus generate a grammatically perfect description of a crystal synthesis that is thermodynamically infeasible because it is a statistically likely sequence of tokens. This issue is compounded by the lack of physical knowledge embedded during training; without constraints derived from thermodynamics or quantum mechanics, models are free to propose physically impossible structures [50].

• Inference-Time Challenges: During text generation, decoding strategies like beam search can amplify small initial errors, leading to a cascade of hallucinations in multi-step reasoning tasks [47] [48]. This is particularly dangerous in domains requiring precise numerical outputs, such as predicting lattice parameters or temperature ranges for synthesis. Moreover, ambiguous or poorly structured prompts can trigger the model to "fill in the gaps" with unverified or invented information [47].

Quantitative Benchmarks: Measuring Hallucination and Performance

Establishing robust benchmarks is critical for quantifying the prevalence of hallucinations and tracking the progress of mitigation strategies. Recent research has introduced several specialized benchmarks for evaluating LLMs in materials science tasks.

Table 2: Performance of LLMs on Materials Science Benchmarks

Benchmark	Core Task	Key Finding	Reference
AtomWorld	Spatial reasoning on CIF files (e.g., structural editing)	Models make frequent errors in structural understanding and spatial reasoning, potentially leading to cumulative errors in subsequent analysis.	[50]
CSLLM Framework	Predict synthesizability of 3D crystal structures	A fine-tuned Synthesizability LLM achieved 98.6% accuracy, significantly outperforming traditional stability screening (74.1% for energy above hull).	[7]
ElemwiseRetro	Predict synthesis precursors for inorganic crystals	The model showed 78.6% top-1 accuracy, outperforming a popularity-based baseline (50.4%), and provided a reliable confidence score.	[49]
CSPBench	Evaluate Crystal Structure Prediction (CSP) algorithms	The performance of current CSP algorithms is far from satisfactory, with most failing to identify structures with the correct space groups.	[51]

The AtomWorld benchmark, for instance, is designed to evaluate fundamental "motor skills" in handling Crystallographic Information Files (CIFs). It tests an LLM's ability to perform actions like adding atoms, moving atoms, rotating atomic groups, and creating supercells. Failures in these basic tasks reveal a model's weakness in spatial reasoning, which is a direct source of hallucination when predicting atomic structures [50].

A Multi-Layered Defense: Protocols for Detecting Hallucinations

Effective detection of hallucinations requires a combination of automated techniques and expert-in-the-loop validation. The following protocols provide a methodology for identifying potential hallucinations in LLM-generated materials data.

Retrieval-Based Fact-Checking

This methodology involves cross-referencing LLM outputs against trusted external knowledge bases.

• Experimental Protocol: Automate queries to curated databases such as the Materials Project [34], the Inorganic Crystal Structure Database (ICSD) [7], and computed property databases from DFT. For any generated crystal structure or claimed property, the system should retrieve the closest matching entry and compute similarity metrics. A significant discrepancy flags a potential hallucination.
• Tools Required: API access to materials databases, semantic similarity models for crystal structures (e.g., using composition and symmetry descriptors), and a threshold for acceptable variance in numerical properties.

Self-Consistency and Uncertainty Quantification

This technique leverages the model's own internal mechanisms to gauge confidence.

• Experimental Protocol: For a single query, generate multiple outputs using stochastic decoding (e.g., varying the temperature parameter). For tasks with a single correct answer, like a synthesis temperature, the variance in the outputs serves as a proxy for uncertainty. For complex tasks like precursor prediction, techniques like Self-Consistency of Chain-of-Thought can identify logical inconsistencies across different reasoning paths [47] [48].
• Tools Required: Access to model log-probabilities, scripts for running multiple inference cycles, and clustering algorithms to analyze output diversity.

Physical Plausibility Checks

This is a domain-specific detection method that applies hard constraints from materials science.

• Experimental Protocol: Implement rule-based filters to check for fundamental physical and chemical laws. This includes:
  • Charge Neutrality: Verify that the sum of oxidation states in a proposed compound equals zero.
  • Interatomic Distance Checks: Ensure that generated crystal structures do not place atoms implausibly close together.
  • Thermodynamic Validation: Use machine learning potentials or high-throughput DFT tools like GNoME [34] to quickly estimate the energy above the convex hull (ΔEhull) of a proposed structure. A high positive value indicates a thermodynamically unstable (and likely hallucinated) material.
• Tools Required: Crystal analysis packages (e.g., pymatgen), fast ML-based property predictors (e.g., M3GNet [51]), and a database of ionic radii.

The following workflow diagram integrates these detection methodologies into a cohesive, sequential process for identifying and flagging hallucinations in LLM-generated materials data.

Mitigation Strategies: Building More Reliable Materials LLMs

Proactively mitigating hallucinations involves architectural, training, and reasoning-based interventions. The following strategies have shown promise in improving the reliability of LLMs for materials science.

Retrieval-Augmented Generation (RAG)

RAG grounds the LLM's generation process by augmenting the prompt with relevant, verifiable information from external knowledge sources [47] [48].

• Implementation Protocol: For a query about a specific material, the system first retrieves the most relevant crystal structures, synthesis protocols, and property data from trusted databases (ICSD, MP). This retrieved context is then prepended to the user's query, forcing the model to base its response on factual evidence. The CSLLM framework effectively employs this principle by using a comprehensive dataset of synthesizable and non-synthesizable crystals to fine-tune its models [7].

Knowledge-Grounded Fine-Tuning

Specialized fine-tuning on high-quality, domain-specific datasets aligns the LLM's knowledge with materials science fundamentals.

• Implementation Protocol: Curate a dataset of verified material structures and synthesis recipes, like the 150,120 crystal structures (70,120 from ICSD, 80,000 non-synthesizable) used for CSLLM [7]. Fine-tune a base LLM on this data using supervised learning. This process sharpens the model's "attention" to domain-specific features, significantly reducing hallucinations related to basic chemistry and synthesis. The ElemwiseRetro model's success is attributed to its training on a text-mined inorganic reaction database, which taught it to use chemically plausible "precursor templates" [49].

Advanced Reasoning and Self-Verification

Structuring the model's reasoning process and incorporating self-checking mechanisms can catch errors before the final output.

• Implementation Protocol: Implement Chain-of-Thought (CoT) prompting that requires the model to reason step-by-step (e.g., "First, identify the source elements in the target composition. Second, select charge-compatible precursors..."). This can be combined with a Self-Consistency check, where multiple reasoning paths are generated and the most consistent answer is selected [48]. For higher-stakes predictions, a more advanced technique like Chain-of-Verification can be used, where the model generates and answers verification questions about its own initial draft [47].

Table 3: The Scientist's Toolkit for Hallucination Mitigation

Tool / Technique	Function	Application Example
Crystallographic Info Files (CIFs)	Standard format for storing crystal structure data; serves as ground truth for retrieval and validation.	Used in the AtomWorld benchmark to test and train LLMs on spatial reasoning tasks [50].
Material String Representation	A simplified, reversible text representation for crystals that integrates lattice, composition, and symmetry data efficiently.	Enabled efficient fine-tuning of the CSLLM framework by providing essential crystal information without CIF redundancy [7].
Graph Neural Networks (GNNs)	ML models that operate on graph-structured data, naturally representing atomic structures and their bonds.	Used by GNoME to predict material stability with high accuracy and by ElemwiseRetro for precursor prediction [34] [49].
Machine Learning Potentials (MLPs)	Fast, surrogate models trained on DFT data that approximate the energy of atomic configurations.	Used in CSP algorithms like GN-OA and AGOX to rapidly screen the stability of predicted crystal structures, flagging hallucinations [51].
Energy Above Convex Hull (ΔEhull)	A thermodynamic metric quantifying the stability of a compound relative to its competing phases.	A primary filter for identifying hallucinated, unstable structures in large-scale discovery efforts like GNoME [34].

The integration of LLMs into the materials discovery pipeline holds immense potential to overcome long-standing bottlenecks in the prediction of synthesizable inorganic crystals. However, realizing this potential requires a systematic and vigilant approach to combating model hallucination. As evidenced by emerging benchmarks like AtomWorld, even advanced models struggle with the fundamental spatial reasoning required for accurate materials modeling [50]. The path forward lies not in seeking a single silver bullet, but in constructing a multi-faceted defense-in-depth strategy. This involves the rigorous application of detection protocols—from retrieval-based fact-checking to physical plausibility checks—coupled with robust mitigation frameworks like Retrieval-Augmented Generation and knowledge-grounded fine-tuning, as demonstrated by the CSLLM and ElemwiseRetro models [7] [49]. By adopting these practices, the materials science community can steer the development of LLMs from sources of creative illusion into reliable, indispensable tools for scientific insight, ultimately accelerating the transition from virtual prediction to real-world synthesis.

The discovery of new inorganic crystalline materials is a fundamental driver of technological advancement, with applications ranging from longer-lived batteries to more efficient solar cells [9]. A central challenge in this field is the vastness of chemical space; while computational methods can screen billions of candidate compositions, only a tiny fraction are synthetically accessible under realistic laboratory conditions [3] [9]. This creates a critical bottleneck in the materials discovery pipeline. The core problem lies in the disconnect between computational stability predictions and experimental synthesizability. Traditional metrics like density functional theory (DFT)-calculated formation energy and distance to the convex hull, while useful, often fail to account for kinetic barriers, finite-temperature effects, and practical synthetic constraints [9] [22]. Consequently, researchers face significant inefficiency, wasting resources on attempting to synthesize materials that are theoretically plausible but experimentally inaccessible.

This whitepaper addresses these fundamental challenges by proposing a framework for confidence scoring—a system that assigns probability metrics to computational predictions to prioritize experimental efforts. By integrating machine learning models that learn from the entire body of previously synthesized materials, these scores provide a calibrated measure of synthesizability, enabling researchers to focus resources on the most promising candidates [3] [22]. This approach represents a paradigm shift from binary classification to probabilistic assessment, offering a more nuanced and efficient strategy for guiding experimental synthesis.

Fundamental Challenges in Predicting Synthesizable Crystals

The Limitation of Thermodynamic Stability

A primary obstacle in computational materials discovery is the overreliance on thermodynamic stability as a proxy for synthesizability. While materials on the convex hull are thermodynamically stable, this condition alone does not guarantee that a material can be synthesized.

• Kinetic and Entropic Factors: Real-world synthesizability is governed by kinetic barriers, entropic effects, and reaction pathways that DFT calculations at zero Kelvin typically overlook [22].
• Metastable Phases: Many experimentally accessible materials are metastable under ambient conditions. For instance, the Materials Project lists 21 SiO₂ structures near the convex hull, yet the common cristobalite phase is not among them, highlighting the limitations of this approach [22].
• False Positives: Accurate regression models for formation energy can still produce high false-positive rates when predictions lie close to the decision boundary, leading to wasted experimental resources [9].

The Data Imbalance Problem

The development of predictive models is severely hampered by the inherent asymmetry in materials data.

• Positive-Unlabeled Learning: While databases like the ICSD catalog synthesized materials, unsuccessful syntheses are rarely reported, creating a scenario with confirmed positive examples but no definitive negative examples [3]. Some materials in databases are labeled as "theoretical" (unsynthesized), but this does not definitively mean they are unsynthesizable [22].
• Sparse Characterization: Vast regions of chemical space remain unexplored, creating an "ignorome" of poorly studied compounds and limiting the coverage of predictive models [52].

The Composition-Structure Gap

Predicting synthesizability is complicated by the interdependent yet distinct roles of composition and crystal structure.

• Unknown Structures: For genuinely novel materials, the crystal structure is typically unknown, necessitating composition-based predictors that are inherently less precise [3].
• Structural Sensitivity: Synthesizability can be highly sensitive to specific structural features, such as local coordination environments and packing motifs, which pure composition-based models miss [22]. A unified model that integrates both composition and structure has been shown to provide more reliable synthesizability assessments [22].

Machine Learning Approaches for Confidence Scoring

Composition-Based Synthesizability Classification

Composition-based models offer a powerful approach for initial large-scale screening due to their applicability even when crystal structures are unknown.

• SynthNN Model: A deep learning model (SynthNN) leverages the entire space of synthesized inorganic chemical compositions from databases like the ICSD [3]. It uses an atom2vec representation to learn optimal chemical descriptors directly from the data, without prior chemical knowledge [3].
• Learned Chemical Principles: Remarkably, without explicit programming, SynthNN learns fundamental chemical principles such as charge-balancing, chemical family relationships, and ionicity from the data distribution of known materials [3].
• Performance Advantage: In benchmarking, SynthNN identified synthesizable materials with 7× higher precision than DFT-calculated formation energies and outperformed a panel of 20 expert material scientists, achieving 1.5× higher precision and completing the task five orders of magnitude faster [3].

Unified Composition and Structure Models

For candidates where structural information is available or can be reliably predicted, integrated models that consider both composition and structure provide superior confidence scores.

• Multi-Encoder Architecture: A state-of-the-art approach uses separate encoders for composition and structure [22]. The composition encoder can be a transformer model, while the structure encoder is typically a graph neural network that operates on the crystal structure graph [22].
• Rank-Average Ensemble: Predictions from both encoders are aggregated using a rank-average ensemble (Borda fusion). This method converts the probability outputs from each model into ranks across the candidate pool and averages these ranks to produce a final, enhanced prioritization score [22]. This technique helps mitigate the limitations of either model in isolation.
• Experimental Validation: This unified approach has been validated experimentally. When applied to screen 4.4 million simulated structures, it identified 24 high-priority candidates, leading to the successful synthesis of 7 out of 16 attempted targets, including one completely novel material [22].

Retrosynthetic Planning with Confidence Estimation

Beyond assessing inherent synthesizability, predicting viable synthesis pathways is crucial. Template-based graph neural networks have been developed for inorganic retrosynthesis.

• ElemwiseRetro Model: This model formulates retrosynthesis by first identifying "source elements" (those that must be provided as precursors) and then attaching appropriate anionic frameworks ("precursor templates") to complete the precursor compounds [20].
• Probability as Confidence: The model calculates a joint probability for sets of precursors, providing a natural confidence score that correlates strongly with prediction accuracy. This score allows researchers to prioritize which predicted recipes to attempt first [20].
• Performance: This model demonstrated a top-1 exact match accuracy of 78.6% and a top-5 accuracy of 96.1%, significantly outperforming a popularity-based baseline [20].

Table 1: Performance Comparison of Machine Learning Models for Materials Discovery

Model Name	Model Type	Input Data	Key Performance Metric	Advantage
SynthNN [3]	Deep Learning (Composition)	Chemical Formula	7x higher precision than DFT formation energy	No structure required; learns chemistry from data
Unified Model [22]	Ensemble (Composition + Structure)	Formula & Crystal Structure	Successfully synthesized 7/16 predicted targets	Integrates multiple signals for higher accuracy
ElemwiseRetro [20]	Graph Neural Network (Retrosynthesis)	Target Composition	78.6% top-1 exact match accuracy	Predicts precursors and provides confidence score
SPaDe-CSP [31]	Workflow (Organic CSP)	Molecular Structure	80% success rate in crystal structure prediction	Reduces generation of unstable structures

A Framework for Experimental Prioritization

Implementing a Confidence-Guided Pipeline

Translating confidence scores into an efficient experimental workflow requires a systematic, staged approach. The following diagram outlines a synthesizability-guided discovery pipeline that integrates the confidence scoring mechanisms discussed.

This pipeline can be operationalized through the following key stages:

• Initial High-Throughput Screening: Apply a composition-based synthesizability model (e.g., SynthNN) to millions of candidate structures from databases like the Materials Project, GNoME, or Alexandria. This rapidly filters out clearly non-viable compositions [3] [22].
• Refined Structural Assessment: For candidates passing the initial screen, employ a unified model that integrates both composition and structural descriptors. This provides a more reliable synthesizability score by accounting for coordination environments and structural motifs [22].
• Rank-Average Ensemble Prioritization: Aggregate scores from different models using a rank-average method (Borda fusion) to create a final, robust priority list. This technique minimizes the impact of model-specific biases and probability miscalibrations [22].
• Retrosynthetic Pathway Prediction: For top-ranked candidates, use a retrosynthesis model (e.g., ElemwiseRetro) to predict viable precursor sets and synthesis conditions, such as calcination temperature. The associated probability scores help prioritize which recipes to attempt first [20] [22].
• Experimental Execution: Execute synthesis attempts based on the predicted recipes using high-throughput automated laboratories, followed by rapid characterization via techniques like X-ray diffraction to verify successful synthesis [22].

Key Metrics for Evaluating Confidence Scores

Implementing this framework requires careful attention to the evaluation metrics used to assess model performance. Common regression metrics like Mean Absolute Error (MAE) can be misleading; a model can have excellent MAE yet high false-positive rates if predictions cluster near the decision boundary [9]. Therefore, the following classification metrics are more appropriate for evaluating confidence scores intended for experimental prioritization:

• Precision-Recall Curves: Especially important for imbalanced datasets where the positive class (synthesizable materials) is rare [3].
• False Positive Rate (FPR) at Thresholds: Critical for understanding how many non-synthesizable materials might pass the filter and consume resources [9].
• Top-k Accuracy: Relevant for retrosynthesis models, measuring whether a valid recipe appears in the top-k predictions [20].
• Area Under the Receiver Operating Characteristic Curve (AUROC) and Area Under the Precision-Recall Curve (AUPRC): Provide aggregate measures of classification performance, with AUPRC being particularly informative for imbalanced data [22].

Table 2: Essential Research Reagents and Computational Tools for Confidence Scoring

Reagent / Tool Category	Specific Examples	Function in Workflow
Materials Databases	Materials Project [22], ICSD [3], GNoME [22], Alexandria [22]	Sources of known and predicted crystal structures for training and screening.
Composition Encoders	MTEncoder transformer [22], atom2vec [3]	Converts chemical formulas into numerical representations for ML models.
Structure Encoders	Graph Neural Networks (GNNs) [22], JMP model [22]	Converts crystal structures (atomic coordinates, bonds) into numerical representations.
Retrosynthesis Models	ElemwiseRetro [20], Retro-Rank-In [22], SyntMTE [22]	Predicts precursor sets and synthesis conditions (e.g., temperature) for a target material.
Validation & Benchmarking	Matbench Discovery [9]	Standardized framework for evaluating model performance on discovery tasks.

The implementation of probability-based confidence scoring represents a critical advancement in the quest to bridge the gap between computational materials prediction and experimental synthesis. By moving beyond traditional thermodynamic metrics and embracing machine learning models that learn the complex, multi-faceted nature of synthesizability from experimental data, researchers can significantly increase the efficiency of materials discovery. The frameworks and models discussed—from composition-based classifiers and unified structure-composition models to retrosynthetic planners with built-in confidence estimates—provide a practical toolkit for prioritizing experimental efforts. As these confidence-scoring methodologies continue to mature and integrate more deeply with high-throughput experimental platforms, they promise to accelerate the discovery and development of next-generation functional materials.

Benchmarking Performance: How New Models Stack Up Against Experts and Traditional Methods

The discovery of novel inorganic crystalline materials is a cornerstone for advancements in various technologies, from energy storage to semiconductors. However, a fundamental challenge persists in predicting whether a computationally designed material is synthesizable—that is, synthetically accessible through current laboratory methods. The traditional paradigm, reliant on human expertise and intuition, is being transformed by artificial intelligence (AI). This guide examines the evolving competition and collaboration between AI and human experts in tackling the synthesizability challenge, providing a technical analysis of their respective capabilities as evidenced by recent, rigorous studies.

Quantitative Performance Comparison

Direct, head-to-head comparisons between AI models and human experts reveal a significant shift in capability. The table below summarizes key performance metrics from recent benchmarking studies.

Table 1: Performance Comparison: AI vs. Human Experts in Material Discovery Tasks

Metric	AI Model (SynthNN)	Best Human Expert	Notes
Precision in Identifying Synthesizable Materials	7x higher than DFT-based formation energy screening [3]	1.5x lower precision than AI [3]	Precision measured against known synthesized materials.
Task Completion Time	Minutes to hours [3]	Weeks to months [3] [53]	AI's speed advantage is multiple orders of magnitude.
Generalization Ability (CSLLM Framework)	98.6% accuracy [17]	Not directly quantified	Accuracy on a balanced dataset of synthesizable/non-synthesizable crystals.
Primary Limitation	Can generate physically implausible structures [54]	Limited by individual experience and domain knowledge [55]	AI's limitation stems from training data; humans are limited by cognitive scope.

These quantitative results demonstrate that AI has surpassed human experts in key aspects of throughput and precision for specific discovery tasks. However, this does not render the human expert obsolete. Instead, it highlights a transition towards a collaborative, "human-in-the-loop" model, where AI handles high-throughput screening and generation, while experts provide critical domain knowledge and feasibility checks [53].

Detailed Experimental Protocols

To understand the performance data, it is essential to examine the underlying methodologies of both AI and human-driven approaches.

AI-Driven Discovery Protocols

• Objective: To train a deep learning model (SynthNN) to classify the synthesizability of inorganic chemical formulas without structural information.
• Data Curation:
  • Positive Samples: Chemical formulas of synthesizable materials were extracted from the Inorganic Crystal Structure Database (ICSD).
  • Negative Samples: Artificially generated chemical formulas that are not present in the ICSD were used as proxies for unsynthesizable materials. The model employs a Positive-Unlabeled (PU) learning approach to account for the possibility that some "unsynthesized" materials may be synthesizable.
• Model Architecture & Training:
  • The model uses an atom2vec framework, which learns an optimal numerical representation (embedding) for each element directly from the distribution of synthesized materials.
  • These embeddings are processed by a neural network optimized using binary cross-entropy loss to predict synthesizability.
• Validation: Performance was benchmarked against a baseline of charge-balancing rules and random guessing, with superior precision and F1-score.

• Objective: To create an end-to-end pipeline that prioritizes predicted crystal structures for experimental synthesis.
• Synthesizability Scoring:
  • A model integrates both compositional (f_c) and structural (f_s) information.
  • Composition is encoded using a fine-tuned transformer (MTEncoder), while crystal structure is encoded using a pretrained graph neural network (JMP).
  • The final synthesizability score is a rank-average ensemble of the probabilities from both models: RankAvg(i) = (1/2N) * Σ [1 + Σ 1(s_m(j) < s_m(i))] for m in {c, s}.
• Synthesis Planning:
  • High-scoring candidates are passed to a precursor-suggestion model (Retro-Rank-In) to identify viable solid-state precursors.
  • A separate model (SyntMTE) predicts the required calcination temperature.
• Experimental Execution: Reactions are balanced, and precursor quantities are computed. Synthesis is executed in a high-throughput laboratory platform, with products characterized by X-ray diffraction (XRD).

• Objective: To predict synthesizability, synthetic methods, and precursors for 3D crystal structures using fine-tuned Large Language Models.
• Data & Representation:
  • A balanced dataset of 70,120 synthesizable structures (from ICSD) and 80,000 non-synthesizable structures (screened via a PU learning model) was created.
  • Crystal structures are converted into a simplified text representation ("material string") that includes essential lattice, composition, and symmetry information.
• Model Fine-Tuning: Three separate LLMs were fine-tuned for distinct tasks:
  • Synthesizability LLM: Binary classification of a structure's synthesizability.
  • Method LLM: Classification of the appropriate synthetic method (e.g., solid-state or solution).
  • Precursor LLM: Identification of suitable solid-state precursors.
• Validation: The framework was tested on theoretical structures from major databases, identifying over 45,000 as synthesizable.

Human Expert-Driven Discovery Protocols

The traditional expert-led approach is less algorithmic and more heuristic, relying on accumulated knowledge and intuition [55].

• Data Curation & Labeling: Experts curate datasets based on personal experience and domain literature. For instance, in identifying topological semimetals, an expert might select a family of "square-net" compounds and label them based on known band structures or chemical analogy [55].
• Descriptor Identification: Experts formulate quantitative descriptors based on physical intuition. An example is the "tolerance factor" for square-net materials, defined as the ratio of the square-net distance to the out-of-plane nearest-neighbor distance (t = d_sq / d_nn) [55].
• Hypothesis Testing: Candidates are selected based on these descriptors, followed by attempts at synthesis and characterization, a process that is iterative and time-consuming.

Workflow Visualization

The following diagram illustrates a modern, AI-integrated materials discovery pipeline, highlighting the complementary roles of AI and human experts.

Synthesizability Guided Discovery Pipeline

The Scientist's Toolkit: Research Reagent Solutions

This section details key computational and experimental resources essential for modern materials discovery workflows.

Table 2: Essential Tools for AI-Accelerated Materials Discovery

Tool / Resource	Type	Primary Function	Relevance to Synthesizability
ICSD (Inorganic Crystal Structure Database) [3] [17]	Database	Repository of experimentally synthesized and characterized inorganic crystal structures.	Serves as the primary source of "positive" data for training and benchmarking AI synthesizability models.
Materials Project, GNoME, Alexandria [22]	Database	Vast collections of DFT-calculated and AI-predicted crystal structures.	Provides the pool of candidate structures that require synthesizability screening to prioritize experimental efforts.
CSLLM (Crystal Synthesis LLM) [17]	AI Model (LLM)	Predicts synthesizability, suggests synthetic methods, and identifies precursors from crystal structure.	Directly addresses the core challenge by providing an end-to-end prediction of synthetic accessibility and pathways.
SynthNN [3]	AI Model (Deep Learning)	Classifies synthesizability of a material based on its chemical composition alone.	Enables rapid screening of vast compositional spaces before committing to structural calculations.
Rank-Average Ensemble Model [22]	AI Model (Ensemble)	Combines compositional and structural model scores for robust synthesizability ranking.	Improves prioritization accuracy over single-model approaches, reducing false positives.
High-Throughput Robotic Synthesis Platform [22]	Experimental	Automates the solid-state synthesis of powdered inorganic samples.	Allows for the rapid experimental validation of AI predictions, closing the discovery loop.
X-ray Diffraction (XRD) [22]	Characterization	Determines the crystal structure of a synthesized powder sample.	The definitive method for verifying if a synthesis attempt successfully produced the target crystal structure.

The "head-to-head" competition between AI and human experts in materials discovery is yielding a definitive outcome: collaboration, not replacement. AI models have demonstrated superior speed and precision in identifying synthesizable candidates from vast chemical spaces, a task where human cognition is a natural bottleneck. However, these models operate within the constraints of their training data and can produce physically implausible suggestions. The human expert's role is evolving from manual screening to that of a crucial validator and integrator, applying irreplaceable domain knowledge to assess real-world feasibility, economic viability, and strategic direction. The most powerful future for materials discovery lies in human-AI synergy, where AI acts as a powerful engine for generation and initial screening, and human intelligence provides the guiding framework of scientific intuition and practical wisdom.

The discovery of new inorganic crystalline materials is a cornerstone of technological advancement, fueling innovations in fields from renewable energy to electronics. However, a formidable bottleneck persists: the vast majority of materials predicted by computational models, even those calculated to be thermodynamically stable, are never successfully synthesized in the laboratory. This critical gap between theoretical prediction and experimental realization represents one of the fundamental challenges in materials science today. The core of the problem lies in accurately predicting synthesizability—whether a material is synthetically accessible through current experimental capabilities, regardless of whether it has been made before. Synthesizability is a complex property influenced not only by thermodynamics but also by kinetic barriers, precursor choices, and specific experimental conditions, factors that traditional stability metrics often fail to capture adequately. This whitepaper provides a quantitative comparison of emerging approaches—specifically Machine Learning (ML) and Large Language Models (LLMs)—against traditional stability metrics for predicting the synthesizability of inorganic crystals, framing this comparison within the broader thesis of overcoming the fundamental challenges in synthesizable materials discovery.

Quantitative Benchmarks: A Comparative Analysis

Recent studies have established rigorous quantitative benchmarks for synthesizability prediction, moving beyond traditional proxies like thermodynamic stability. The table below summarizes key performance metrics across different prediction paradigms.

Table 1: Quantitative Benchmarks for Synthesizability Prediction Accuracy

Prediction Method	Reported Accuracy	Key Metric	Dataset & Context
CSLLM (LLM-based)	98.6% [7]	Overall Accuracy	Balanced dataset of 70,120 synthesizable (ICSD) and 80,000 non-synthesizable structures [7]
PU Learning (ML-based)	87.9% [7]	Overall Accuracy	3D crystal structures; pre-trained model used to filter non-synthesizable examples [7]
Teacher-Student NN (ML-based)	92.9% [7]	Overall Accuracy	Improvement on previous PU learning models for 3D crystals [7]
Traditional (Kinetic Stability)	82.2% [7]	Overall Accuracy	Screening based on phonon spectrum (lowest frequency ≥ -0.1 THz) [7]
Traditional (Thermodynamic Stability)	74.1% [7]	Overall Accuracy	Screening based on energy above hull (≥ 0.1 eV/atom) [7]
SynthNN (ML-based, composition-only)	7x higher precision than DFT [3]	Precision	Trained on ICSD data with artificially generated unsynthesized materials; outperformed human experts [3]
Charge-Balancing Heuristic	~37% of known materials are charge-balanced [3]	Coverage	Applied to all synthesized inorganic materials in ICSD [3]

The data reveals a clear performance hierarchy. LLM-based approaches, particularly the Crystal Synthesis Large Language Model (CSLLM), currently set the state-of-the-art, significantly outperforming both traditional methods and earlier ML models. It is critical to note that these accuracy metrics are highly dependent on the dataset and the specific definition of "non-synthesizable" used for training and testing. For instance, the high accuracy of the CSLLM was achieved on a balanced and curated dataset where non-synthesizable examples were identified using a pre-trained PU learning model to screen over 1.4 million theoretical structures [7].

Methodologies and Experimental Protocols

The superior performance of modern synthesizability predictors is rooted in their sophisticated data construction and model architectures. Below, we detail the core methodologies for the leading approaches.

LLM-Based Framework (CSLLM)

The Crystal Synthesis Large Language Models framework represents a paradigm shift by treating crystal structures as text sequences.

1. Data Curation:

• Positive Data: 70,120 experimentally synthesized, ordered crystal structures were meticulously selected from the Inorganic Crystal Structure Database (ICSD). Structures were filtered to a maximum of 40 atoms and 7 different elements [7].
• Negative Data: A key challenge is obtaining reliable non-synthesizable examples. The CSLLM protocol used a pre-trained Positive-Unlabeled (PU) learning model to calculate a "CLscore" for 1.4 million theoretical structures from major databases (Materials Project, OQMD, JARVIS). The 80,000 structures with the lowest CLscores (<0.1) were selected as high-confidence negative examples, creating a balanced dataset [7].

2. Text Representation - "Material String": To fine-tune LLMs, a compact text representation for crystals was developed. This "material string" condenses essential crystallographic information by leveraging symmetry, avoiding the redundancy of CIF or POSCAR files. The format is: SP | a, b, c, α, β, γ | (AS1-WS1[WP1-O1], AS2-WS2[WP2-O2], ...) [7] Where SP is the space group number, a, b, c, α, β, γ are lattice parameters, and the tuple contains atomic species (AS), Wyckoff site symbols (WS), Wyckoff position coordinates (WP), and occupation (O).

3. Model Fine-Tuning: The framework employs three specialized LLMs fine-tuned on this data:

• Synthesizability LLM: A binary classifier predicting if a structure is synthesizable.
• Method LLM: Classifies the likely synthetic pathway (e.g., solid-state or solution).
• Precursor LLM: Identifies suitable precursor compounds for synthesis [7].

Machine Learning (ML) Approaches

1. Positive-Unlabeled (PU) Learning: This semi-supervised approach directly addresses the lack of confirmed negative data. It treats the entire set of unsynthesized materials as "unlabeled," which contains a mix of synthesizable and non-synthesizable compounds. The model is trained to identify the known positives (ICSD) and then probabilistically reweights the unlabeled examples to learn the characteristics of negatives, effectively learning synthesizability from incomplete data [3].

2. Structure-Derivation with Wyckoff Sampling: Some ML frameworks shift from random structure search to a more targeted approach. This method involves:

• Prototype Database Construction: Deriving symmetry-maximized prototype structures from known synthesized materials in databases like the Materials Project [1].
• Group-Subgroup Relations: Systematically generating candidate structures by applying symmetry-reduction chains to these prototypes, ensuring the sampled structures retain spatial arrangements of experimentally realized materials [1].
• Wyckoff Filtering: Classifying generated structures into subspaces based on their Wyckoff encoding and using an ML model to filter for the most promising, synthesizable-rich subspaces before costly computational relaxation [1].

Traditional Stability Metrics

1. Thermodynamic Stability (Formation Energy): This method uses Density Functional Theory (DFT) to calculate a material's formation energy. The energy above the convex hull (ΔE(h)) is the standard metric; it represents the energy difference between the material and its most stable decomposition products into other phases. A ΔE(h) ~ 0 eV/atom indicates thermodynamic stability, but this is a strict criterion that filters out many synthesizable metastable materials [3] [7].

2. Kinetic Stability (Phonon Spectrum): This assesses dynamic stability by computing the phonon dispersion of a crystal structure. The presence of imaginary frequencies (negative values in THz) indicates a saddle point on the potential energy surface, suggesting the structure is unstable to atomic displacements. However, some materials with imaginary frequencies can still be synthesized, making this an imperfect predictor [7].

Workflow and Signaling Pathways

The following diagram illustrates the logical workflow and key decision points for a synthesizability-driven crystal structure prediction (CSP) framework, integrating the methodologies discussed above.

The Scientist's Toolkit: Essential Research Reagents

The experimental and computational research in this field relies on several key "reagents"—databases, software, and models. The following table details these essential components.

Table 2: Key Research Resources for Synthesizability Prediction

Resource Name	Type	Primary Function	Relevance to Synthesizability
ICSD [3] [7]	Database	Repository of experimentally synthesized inorganic crystal structures.	Serves as the primary source of confirmed "positive" data for training and benchmarking models.
Materials Project (MP) [1] [7]	Database	Repository of computationally predicted and characterized materials.	A major source of "unlabeled" or candidate structures; used for generating negative examples and testing.
PU Learning Model [7]	Algorithm/Method	Semi-supervised learning to learn from positive and unlabeled data.	Core to many ML approaches for handling the lack of confirmed negative data.
Wyckoff Representation [1] [56]	Structural Descriptor	A symmetry-aware representation of crystal structures using Wyckoff positions.	Enables efficient, symmetry-compliant generation and filtering of candidate structures, improving search efficiency.
Material String [7]	Data Format	A compact text representation of crystal structure for LLM processing.	Allows LLMs to be fine-tuned on crystallographic data, bridging the gap between structural chemistry and natural language processing.
DFT (VASP, etc.) [56]	Computational Tool	First-principles calculation of formation energy and phonon spectra.	Provides the traditional stability metrics (ΔE(_h), phonons) used as baselines for comparison.
Universal Interatomic Potentials [57]	ML Model	Machine-learned potential for rapid energy and force evaluation.	Used for fast pre-screening and relaxation of generated structures before final DFT validation.

The quantitative benchmarks and methodologies presented herein unequivocally demonstrate a significant evolution in the ability to predict the synthesizability of inorganic crystals. While traditional stability metrics provide a foundational baseline, they are insufficient alone, with accuracy plateauing around 74-82% [7]. Machine learning models, particularly those employing positive-unlabeled learning, marked a substantial improvement, pushing accuracy to nearly 93% [7]. The most transformative advance, however, comes from Large Language Models fine-tuned on crystallographic data. The CSLLM framework's 98.6% accuracy establishes a new state-of-the-art, showcasing the power of reformulating crystal structures as a language problem [7]. This progression is a critical response to the fundamental challenge in materials discovery: closing the gap between computational prediction and experimental realization. By moving beyond a purely energy-based paradigm to a data-driven, synthesis-aware one, these modern tools are forging a more reliable and efficient pathway for transforming theoretical candidate materials into tangible, laboratory-synthesized realities.

The accelerating use of machine learning (ML) for computational materials discovery has unveiled a critical bottleneck: the challenge of model generalization. While ML models can rapidly screen millions of hypothetical crystals, their true utility depends on reliably predicting the synthesizability of structures that are compositionally and structurally distinct from those in their training data. This challenge is fundamental to the broader mission of predicting synthesizable inorganic crystals, as models that fail to generalize beyond their training distribution can misdirect experimental resources toward theoretically appealing but practically inaccessible materials [58]. The core of this problem lies in the fact that models are typically trained on existing experimental databases, which represent only a tiny, potentially biased fraction of the vast chemical space [17]. Consequently, validating performance on novel and complex structures through rigorous generalization tests has become an essential discipline within materials informatics.

This whitepaper provides a comprehensive technical guide to current methodologies for assessing the generalization capability of synthesizability prediction models. We synthesize recent advances from leading research efforts, present standardized quantitative evaluation frameworks, and detail experimental protocols for stress-testing models against structurally complex and compositionally novel materials. By establishing robust validation standards, the field can enhance the reliability of computational predictions and accelerate the experimental realization of novel functional materials.

Computational Frameworks for Synthesizability Prediction

Current approaches for predicting material synthesizability primarily fall into three categories: composition-based, structure-based, and hybrid models. Composition-based models, such as SynthNN, operate solely on chemical formulas and leverage learned representations of elements and their stoichiometries to predict synthesizability [3]. These models benefit from applicability across vast compositional spaces but cannot distinguish between different polymorphs of the same composition. Structure-based models require full crystallographic information (lattice parameters, atomic coordinates, space groups) as input. The Crystal Synthesis Large Language Models (CSLLM) framework represents a recent advancement in this category, achieving high accuracy by converting crystal structures into specialized text representations processed by fine-tuned large language models [17]. Hybrid models integrate both compositional and structural descriptors; for example, some pipelines combine compositional transformers with graph neural networks operating on crystal structures, then aggregate predictions through rank-average ensembling to enhance robustness [22].

A significant challenge in training these models is the inherent asymmetry in materials data: while synthesizable examples exist in curated databases, definitively non-synthesizable examples are scarce and must be artificially generated or identified through semi-supervised techniques [3] [59]. This has led to the adoption of Positive-Unlabeled (PU) learning frameworks, where models are trained on known synthesized materials (positives) and large sets of unlabeled candidates, with the latter probabilistically reweighted based on their likelihood of being synthesizable [3] [17].

Table 1: Overview of Synthesizability Prediction Model Types

Model Type	Key Input Features	Representative Examples	Strengths	Limitations
Composition-Based	Elemental stoichiometry	SynthNN [3]	Computationally lightweight; screens vast compositional space	Cannot distinguish polymorphs
Structure-Based	Crystal structure (lattice, atomic coordinates, symmetry)	CSLLM [17]	Accounts for structural stability and packing	Requires full structural information
Hybrid	Both composition and crystal structure	Rank-average ensemble models [22]	Leverages complementary signals from composition and structure	Increased complexity

Quantitative Benchmarks and Performance Metrics

Rigorous quantification of model performance on held-out test sets provides the foundation for generalization assessment. Recent state-of-the-art models demonstrate impressive accuracy on standard benchmarks, but these metrics must be interpreted with caution due to potential data biases.

The CSLLM framework reports a remarkable 98.6% accuracy on its test set, significantly outperforming traditional thermodynamic stability screening based on formation energy (74.1% accuracy) and kinetic stability assessment via phonon spectrum analysis (82.2% accuracy) [17]. Similarly, composition-based models like SynthNN have demonstrated 7× higher precision in identifying synthesizable materials compared to density functional theory (DFT)-calculated formation energies alone [3]. In competitive benchmarking against human experts, SynthNN achieved 1.5× higher precision in material discovery tasks while completing the task five orders of magnitude faster [3].

Other approaches using semi-supervised learning for stoichiometry synthesizability prediction report true positive rates of 83.4% with an estimated precision of 83.6% [59]. Meanwhile, hybrid models integrating composition and structure have successfully guided experimental campaigns, resulting in the synthesis of 7 out of 16 targeted compounds, demonstrating a tangible real-world success rate of 44% for computationally predicted candidates [22].

Table 2: Performance Benchmarks of Synthesizability Prediction Methods

Evaluation Method	SynthNN (Composition) [3]	CSLLM (Structure) [17]	Semi-Supervised Learning [59]	Hybrid Model [22]
Accuracy	Not specified	98.6%	Not specified	Not specified
Precision	7× higher than DFT	Not specified	83.6% (estimated)	44% experimental success
True Positive Rate	Not specified	Not specified	83.4%	Not specified
Comparison Baseline	DFT formation energy	Thermodynamic (74.1%) and kinetic (82.2%) stability	Not specified	Experimental validation

Methodologies for Generalization Testing

Structural Complexity Tests

A powerful approach to stress-test generalization involves evaluating model performance on crystals with increasing structural complexity, particularly those with large unit cells containing many atoms. The CSLLM framework demonstrated exceptional generalization using this method, achieving 97.9% accuracy on structures with complexity considerably exceeding that of its training data [17]. This test is implemented by:

• Curating a complexity-graded dataset: Extract structures from databases like the Inorganic Crystal Structure Database (ICSD) with varying numbers of atoms per unit cell, ensuring representation of simple to highly complex configurations.
• Stratified performance analysis: Partition the test set by complexity metrics (atoms per unit cell, number of unique elements, space group symmetry) and evaluate accuracy separately for each complexity tier.
• Out-of-distribution detection: Monitor performance degradation as structural complexity increases beyond the training distribution median, quantifying the model's robustness to unfamiliar structural environments.

Compositional Novelty Assessments

Compositional generalization tests evaluate model performance on chemical formulas containing element combinations poorly represented in training data. The protocol involves:

• Leave-out-element cross-validation: Systematically exclude all compounds containing specific elements during training, then test the model on compounds containing those held-out elements.
• Ternary and quaternary compound validation: Focus testing on higher-order compounds (e.g., quaternary oxides) that present more complex bonding environments and charge-balancing constraints than the binary compounds typically well-represented in training sets.
• Phase map construction: Generate synthesizability predictions across continuous compositional spaces (e.g., metal oxide systems) and validate predictions against known experimental phase diagrams to verify the model captures chemically intuitive boundaries [59].

Temporal Hold-Out Validation

This method assesses a model's ability to predict materials discovered after its training data was collected, simulating real-world discovery scenarios:

• Time-splitting: Train models exclusively on structures reported before a specific cutoff date (e.g., 2015).
• Future prediction: Evaluate performance on structures discovered after the cutoff date, quantifying the model's capacity to generalize to truly novel discoveries rather than merely interpolating within existing knowledge.
• Progressive evaluation: Repeat the process with multiple temporal cutoffs to establish trends in model performance as a function of temporal displacement.

Experimental Validation and Closed-Loop Testing

The most rigorous generalization test involves guiding actual laboratory synthesis efforts, creating a closed-loop validation pipeline as implemented by Prein et al. [22]:

• High-throughput computational screening: Apply the synthesizability model to millions of candidate structures from computational databases like the Materials Project, GNoME, and Alexandria.
• Candidate prioritization: Select top-ranked candidates based on synthesizability scores while applying practical filters (excluding toxic elements, rare/expensive components).
• Retrosynthetic analysis: Employ precursor-suggestion models (e.g., Retro-Rank-In) to identify viable solid-state precursors and predict calcination temperatures.
• Automated synthesis and characterization: Execute high-throughput synthesis using automated laboratory platforms and characterize products via X-ray diffraction to verify target phase formation.

This end-to-end validation provides the most credible assessment of real-world utility, with successful demonstrations yielding experimental synthesis of novel compounds predicted by the models [22] [59].

Visualization of Experimental Validation Workflow

The following diagram illustrates the comprehensive experimental workflow for validating synthesizability model predictions through laboratory synthesis, as described in Section 4.4:

Diagram 1: Experimental validation workflow for synthesizability models.

The Scientist's Toolkit: Essential Research Reagents

Implementation of generalization tests requires specific computational and experimental resources. The following table details key components of the research infrastructure for synthesizability prediction and validation:

Table 3: Essential Research Reagents for Synthesizability Prediction Research

Research Reagent	Type	Function in Generalization Testing	Example Sources
ICSD (Inorganic Crystal Structure Database)	Data Resource	Provides experimentally synthesized crystal structures for training and benchmarking positive examples	ICSD [3] [17]
Materials Project	Data Resource	Supplies computationally generated candidate structures for creating negative examples and screening pools	Materials Project [17] [22]
PU Learning Model	Computational Algorithm	Implements positive-unlabeled learning to handle lack of confirmed negative examples	CLscore model [17]
CSLLM Framework	Software Tool	Predicts synthesizability, synthetic methods, and precursors for crystal structures	CSLLM [17]
Retro-Rank-In	Computational Algorithm	Suggests viable solid-state precursors for target compounds	Retro-Rank-In [22]
High-Throughput Synthesis Platform	Experimental System	Enables rapid experimental validation of computational predictions	Automated lab platforms [22]

Discussion: Remaining Challenges and Future Directions

Despite significant advances, important challenges persist in generalization testing for synthesizability prediction. Data bias remains a fundamental concern, as models trained on heterogeneous datasets (mixing experimental and computational sources) may learn spurious correlations that limit real-world applicability [58]. The molecular assembly problem for non-polymeric crystals presents particular difficulties, as current benchmarks may not adequately capture the permutation invariance of identical molecular units in crystal structures [60]. Additionally, while thermodynamic stability metrics like formation energy and energy above the convex hull provide useful references, they remain imperfect proxies for synthesizability, with many metastable structures being synthesizable and numerous thermodynamically stable structures remaining elusive [17] [22].

Future progress will likely come from several directions: improved domain-specific loss functions that better capture physical principles of crystal packing [60], enhanced data collection strategies that mitigate sampling bias [58], and more sophisticated multi-task learning frameworks that simultaneously predict synthesizability, synthesis pathways, and suitable precursors [17]. The development of standardized benchmarks with stratified complexity metrics will enable more systematic comparison of generalization capabilities across different modeling approaches. Furthermore, increased emphasis on experimental validation through closed-loop discovery pipelines will provide the ultimate test of model utility in real materials discovery campaigns.

As the field advances, the rigorous generalization testing methodologies outlined in this whitepaper will play an increasingly critical role in ensuring that computational synthesizability predictions translate successfully from theoretical models to laboratory realities, ultimately accelerating the discovery of novel functional materials for energy, electronics, and other transformative technologies.

The accurate prediction of precursor materials represents a fundamental challenge in the design of synthesizable crystals, bridging the fields of organic chemistry and inorganic materials science. In retrosynthesis planning, a core task is to work backward from a target compound to deduce a set of simpler precursor compounds that can feasibly synthesize it [28]. The "Top-k exact match accuracy" has emerged as a critical benchmark for evaluating model performance in this domain, measuring the probability that the model's ranked list of precursor suggestions contains the exact, historically verified set of starting materials within the top-k recommendations [20] [61]. This metric provides a rigorous standard for assessing the practical utility of retrosynthesis algorithms, yet its interpretation varies significantly between the distinct challenges of organic molecule synthesis and inorganic materials formation. This technical guide examines the state-of-the-art in precursor prediction accuracy, detailing the experimental methodologies, performance benchmarks, and underlying architectures that define current capabilities and limitations in this rapidly evolving field.

Performance Benchmarks: Top-k Accuracy Across Model Types

Performance on Standardized Organic Chemistry Benchmarks

Retrosynthesis models for organic chemistry are primarily evaluated on the USPTO-50k dataset, which contains approximately 50,000 atom-mapped reaction examples [62] [61]. The table below summarizes the Top-k exact match accuracy of contemporary models on this benchmark, demonstrating the progression toward higher prediction accuracy.

Table 1: Top-k exact match accuracy (%) of retrosynthesis models on the USPTO-50k dataset

Model	Type	Top-1	Top-3	Top-5	Top-10
RSGPT [63]	Template-free (LLM)	63.4	-	-	-
EditRetro [64]	Template-free (String Editing)	60.8	-	-	-
TorchDrug (Given Reaction Class) [65]	Not Specified	63.9	85.2	90.4	93.8
TorchDrug (Unknown Reaction Class) [65]	Not Specified	43.8	67.7	74.8	82.2
Graph2Edits [62]	Semi-template-based (Graph Editing)	55.1	-	-	-

These results highlight several key trends. First, the highest-performing models now achieve Top-1 accuracies exceeding 60%, representing significant progress in the field [64] [63]. Second, performance improves substantially when reaction class information is provided, as evidenced by the nearly 20-point difference in Top-1 accuracy for the TorchDrug model [65]. This underscores the importance of chemical context in accurate precursor prediction. Finally, the diversity of architectural approaches—from large language models (RSGPT) to string-editing (EditRetro) and graph-editing (Graph2Edits) methods—demonstrates multiple viable pathways for addressing the retrosynthesis challenge.

Performance on Inorganic Materials Synthesis

For inorganic materials synthesis, the evaluation metrics and datasets differ substantially from organic chemistry. The following table presents the Top-k exact match accuracy for inorganic retrosynthesis models, which predict solid-state precursor sets for target inorganic compositions.

Table 2: Top-k exact match accuracy (%) of retrosynthesis models for inorganic materials

Model	Top-1	Top-2	Top-3	Top-4	Top-5
ElemwiseRetro (RandSplit) [20]	78.6	87.7	92.9	94.6	96.1
ElemwiseRetro (TimeSplit) [20]	80.4	89.4	92.9	94.3	95.8
Popularity Baseline [20]	50.4	70.5	75.1	77.6	79.2

The notably higher accuracy scores for inorganic retrosynthesis reflect fundamental differences in the problem domain. Unlike organic synthesis with its vast potential reaction pathways, solid-state inorganic synthesis typically utilizes a finite set of commercially available precursors, simplifying the prediction task [20]. The ElemwiseRetro model significantly outperforms the popularity-based baseline, demonstrating its ability to learn meaningful chemical relationships beyond simple frequency statistics [20]. The TimeSplit results are particularly noteworthy, showing the model's ability to generalize to materials synthesized after its training period, a crucial capability for predicting precursors for novel materials [20].

Methodological Approaches and Experimental Protocols

Core Architectural Paradigms for Retrosynthesis

Template-Free Organic Retrosynthesis

Contemporary template-free approaches have moved beyond simple sequence-to-sequence translation, incorporating more sophisticated editing-based strategies. EditRetro reframes retrosynthesis as a molecular string editing task, iteratively refining target molecule strings to generate precursor compounds [64]. The model employs a fragment-based generative editing approach using explicit sequence editing operations (Levenshtein operations) including repositioning, placeholder insertion, and token insertion [64]. This methodology leverages the significant structural overlap between reactants and products characteristic of most chemical reactions. The experimental protocol involves training on standardized SMILES representations of product-reactant pairs, with the model learning to predict a sequence of edit operations that transform the product into its precursors [64].

For inference, EditRetro incorporates reposition sampling and sequence augmentation to enhance prediction diversity [64]. Sequence augmentation creates variants of canonical molecular SMILES by randomly selecting starting atoms and enumeration directions, enabling diverse editing pathways from product strings to reactants [64]. This approach demonstrates how explicit incorporation of chemical intuition—recognizing that reactions typically involve local molecular changes—can drive significant performance improvements, achieving a top-1 accuracy of 60.8% on USPTO-50k [64].

Semi-Template-Based Organic Retrosynthesis

Semi-template-based methods strike a balance between template-based and template-free approaches. Graph2Edits exemplifies this paradigm, implementing an end-to-end graph editing framework inspired by arrow-pushing formalism in chemical reaction mechanisms [62]. The model predicts a sequence of graph edits—including bond changes and functional group additions/removals—that transform the product graph into reactant graphs [62].

The experimental workflow for Graph2Edits involves several key stages. First, the model represents the product molecule as a graph, with atoms as nodes and bonds as edges [62]. A graph neural network then processes this representation to predict a sequence of edits [62]. These edits are applied sequentially to generate transformation intermediates and final reactants [62]. This approach combines the advantages of both template-based and template-free methods: it provides the interpretability of explicit structural transformations while maintaining the generalization capability of template-free systems [62]. On the USPTO-50k dataset, this methodology achieves a top-1 accuracy of 55.1% [62].

Inorganic Retrosynthesis Frameworks

Inorganic retrosynthesis requires fundamentally different approaches due to the distinct nature of solid-state materials synthesis. The ElemwiseRetro framework formulates the problem through element-wise decomposition [20]. The approach first categorizes elements in the target composition as either "source elements" (must be provided as reaction precursors) or "non-source elements" (can come from or leave reaction environments) [20]. For each source element, the model selects appropriate precursor templates from a library of common solid-state precursors [20].

The experimental protocol involves several stages. The target composition is encoded as a graph with node features derived from pretrained representations of inorganic compounds [20]. The model then applies a source element mask to identify which elements must be provided by precursors [20]. A precursor classifier predicts the specific precursor compound for each source element using the formulated template library [20]. Finally, the model calculates the joint probability of the complete precursor set, enabling ranking of multiple synthesis recipes by confidence [20]. This methodology achieves a remarkable 78.6% top-1 exact match accuracy on inorganic synthesis datasets [20].

Table 3: Core methodological differences between organic and inorganic retrosynthesis approaches

Aspect	Organic Retrosynthesis	Inorganic Retrosynthesis
Primary Representation	SMILES strings [64] or Molecular graphs [62]	Elemental compositions & precursor templates [20]
Key Challenge	Handling diverse reaction mechanisms & functional group compatibility	Selecting commercially available precursors that provide required elements
Typical Output	Specific reactant molecules	Sets of precursor compounds
Evaluation Metric	Exact match of predicted reactants [61]	Exact match of precursor sets [20]
Data Source	USPTO-50k, USPTO-FULL [64] [63]	Text-mined inorganic synthesis databases [20]

Training Methodologies and Data Curation

Data Preparation Protocols

Standardized data preparation is crucial for reproducible retrosynthesis model evaluation. For organic chemistry, the USPTO-50k dataset serves as the primary benchmark, containing 50,016 reactions with correct atom-mapping classified into 10 reaction types [62]. The standard data split allocates 40,000 reactions for training, 5,000 for validation, and 5,000 for testing [62]. To prevent information leakage, researchers canonicalize product SMILES and re-assign mapping numbers to reactant atoms following established protocols [62].

For inorganic retrosynthesis, data is typically curated from sources like the Cambridge Structural Database or text-mined synthesis literature [20]. The ElemwiseRetro study used 13,477 curated inorganic retrosynthetic datasets to extract 60 precursor templates [20]. Data filtering criteria often include lattice parameter ranges (2 ≤ a, b, c ≤ 50 Å; 60 ≤ α, β, γ ≤ 120°) to exclude extreme outliers and ensure data quality [31].

Advanced Training Paradigms

Recent approaches have incorporated training strategies from large language models to address data scarcity. RSGPT employs a three-stage training process: pre-training on massive synthetic data, reinforcement learning from AI feedback (RLAIF), and task-specific fine-tuning [63]. The model generates over 10 billion synthetic reaction data points using template-based algorithms, then pre-trains on this expanded corpus to acquire comprehensive chemical knowledge [63]. During RLAIF, the model generates reactants and templates for given products, with RDChiral validating the rationality of outputs and providing reward signals to refine the model's understanding of the relationships between products, reactants, and templates [63]. This innovative approach achieves a state-of-the-art top-1 accuracy of 63.4% on USPTO-50k [63].

Experimental Workflows and Research Tools

Key Experimental Workflows

The following diagram illustrates the core workflow for edit-based retrosynthesis prediction, as implemented in the EditRetro model:

Edit-Based Retrosynthesis Workflow

For inorganic materials, the precursor prediction workflow follows a distinctly different pathway, as illustrated in the following diagram:

Inorganic Precursor Prediction Workflow

Essential Research Reagents and Computational Tools

Table 4: Key research reagents and computational tools for retrosynthesis experiments

Tool/Resource	Type	Primary Function	Application Context
USPTO-50k [62]	Dataset	Benchmark dataset with ~50k reactions	Organic retrosynthesis evaluation
RDChiral [63]	Algorithm	Template extraction & reaction validation	Synthetic data generation & reaction reasoning
SMILES [64]	Representation	String-based molecular encoding	Organic molecule representation
CIF Files [66]	Representation	Crystallographic information file format	Inorganic crystal structure representation
Graph Neural Networks [62]	Architecture	Graph-structured data processing	Molecular graph representation learning
Transformer Models [63]	Architecture	Sequence-to-sequence prediction	Template-free retrosynthesis
Monte Carlo Tree Search [66]	Algorithm	Search and optimization	Multi-step synthesis planning

The pursuit of higher Top-k exact match accuracy continues to drive innovation in retrosynthesis methodology, with contemporary models achieving remarkable performance through diverse architectural strategies. The progression from template-based to template-free and editing-based approaches has yielded models capable of predicting organic synthesis precursors with over 60% Top-1 accuracy and inorganic solid-state precursors with nearly 80% Top-1 accuracy. These advances stem from sophisticated computational frameworks that incorporate chemical intuition—recognizing the local nature of molecular transformations in organic chemistry and the template-driven precursor selection in inorganic solid-state synthesis. As the field evolves, the integration of large-scale synthetic data generation, reinforcement learning from AI feedback, and more nuanced evaluation metrics promises to further bridge the gap between computational prediction and experimental synthesis, ultimately accelerating the discovery and development of novel functional materials across both organic and inorganic domains.

The discovery of new functional materials is a critical driver of technological progress in areas such as energy storage, carbon capture, and semiconductor design. However, the traditional materials discovery process, reliant on human intuition and experimentation, creates long iteration cycles that fundamentally limit the pace of innovation. The core challenge lies in the vastness of chemical space; while approximately 10^5 material combinations have been tested experimentally and ~10^7 have been simulated, upwards of 10^10 possible quaternary materials are allowed by electronegativity and charge-balancing rules [9]. This explorable space grows even larger for quinternaries and beyond, leaving immense regions of potentially useful materials uncharted.

At the heart of this challenge is the problem of predicting synthesizable inorganic crystals—materials that are not only thermodynamically stable but also capable of being realized in laboratory conditions. The disconnect between thermodynamic stability (as computed by density functional theory) and actual synthesizability represents a critical bottleneck. While high-throughput computational screening has expanded our reach, it remains fundamentally limited by the number of known materials, accessing only a tiny fraction of potentially stable inorganic compounds [67]. This whitepaper assesses the current readiness of artificial intelligence to overcome these fundamental challenges through integrated autonomous discovery workflows, examining recent technical advances, performance benchmarks, and the practical frameworks needed for real-world implementation.

Technical Foundations of AI-Driven Materials Discovery

Generative Architectures for Inverse Design

The paradigm of materials discovery is shifting from high-throughput screening to AI-driven inverse design, where desired property constraints directly inform the generation of candidate materials. Several architectural approaches have emerged as particularly powerful for this task:

Diffusion models have demonstrated remarkable success in generating stable, diverse inorganic materials across the periodic table. MatterGen, a diffusion-based generative model, employs a customized diffusion process that generates crystal structures by gradually refining atom types, coordinates, and the periodic lattice. Its corruption process respects the unique periodic structure and symmetries of crystalline materials, using a wrapped Normal distribution for coordinate diffusion that approaches a uniform distribution at the noisy limit, and a symmetric form for lattice diffusion that approaches a cubic lattice with average atomic density from training data [67]. This approach more than doubles the percentage of generated stable, unique, and new (SUN) materials compared to previous state-of-the-art methods and produces structures that are more than ten times closer to their DFT-relaxed structures [67].

Adapter modules enable flexible conditioning of generative models on desired property constraints. These tunable components are injected into each layer of a base model to alter its output depending on given property labels, allowing for effective fine-tuning even when labeled datasets are small compared to unlabeled structure datasets [67]. When combined with classifier-free guidance, this approach enables steering generation toward target chemical compositions, symmetry requirements, and scalar properties such as magnetic density.

Cross-modal contrastive learning bridges textual descriptions with structural data to inform generative processes. The Chemeleon model employs Crystal CLIP, a framework that aligns text embedding vectors from a transformer-based encoder with graph embeddings from equivariant graph neural networks (GNNs) through contrastive learning [27]. This alignment is achieved by maximizing cosine similarity for positive pairs (graph embeddings and their corresponding textual descriptions from the same crystal structure) while minimizing similarity for negative pairs, creating a shared latent space where semantically similar crystals and texts are proximate [27].

Machine Learning Potentials for Accelerated Relaxation

The computational bottleneck of density functional theory (DFT) relaxation represents a significant challenge in traditional materials discovery, with DFT demanding up to 45% of core hours at supercomputing facilities [9]. Machine learning potentials have emerged as a transformative solution:

Universal interatomic potentials (UIPs) trained on diverse DFT datasets now cover 90 or more elements in the periodic table, enabling accurate energy and force predictions at a fraction of the computational cost [9]. Benchmark studies demonstrate that UIPs surpass all other methodologies in both accuracy and robustness for pre-screening thermodynamically stable hypothetical materials [9].

Neural network potentials (NNPs) achieve near-DFT-level accuracy while dramatically accelerating structural relaxation. For organic crystals, pre-trained base models such as PFP and ANI have demonstrated efficacy that can surpass quantum chemical methods in accuracy for certain applications [31]. These potentials can be further fine-tuned for specific systems through additional training, making them highly versatile for specialized discovery campaigns.

Performance Benchmarks: Quantitative Assessment of AI Readiness

Rigorous benchmarking is essential for assessing the practical readiness of AI approaches for autonomous discovery. Recent comprehensive evaluations provide critical insights into current capabilities and limitations.

Table 1: Performance Comparison of Generative Models for Inorganic Materials

Model	Architecture	SUN Materials Rate	Average RMSD to DFT (Å)	Novelty Rate	Key Innovation
MatterGen	Diffusion	75%	<0.076	61%	Adapter modules for property conditioning [67]
CDVAE	Variational Autoencoder	~30%	~0.8	-	Early deep learning approach [67]
DiffCSP	Diffusion	~35%	~0.7	-	Specialized for crystal structure prediction [67]
Chemeleon	Text-guided Diffusion	-	-	-	Cross-modal contrastive learning [27]

Table 2: CSP Algorithm Performance Benchmark (CSPBench Evaluation) [51]

Algorithm Category	Representative Methods	Success Rate	Computational Efficiency	Key Limitation
Template-based CSP	TCSP, CSPML	High for similar templates	High	Limited to known structural motifs
DFT-based Global Search	CALYPSO, USPEX	Moderate	Low (DFT-bound)	Extreme computational cost
ML Potential-based Search	GNOA, ParetoCSP	Competitive with DFT	Medium	Dependent on potential quality
Random Search + MLP	AGOX with M3GNet	Moderate	Medium	Less directed exploration

The benchmark results reveal several critical insights. First, generative models have achieved significant improvements in generating stable materials, with MatterGen producing 75% of structures within 0.1 eV per atom above the convex hull of a combined reference dataset [67]. Second, the structural quality of generated materials has dramatically improved, with 95% of MatterGen structures having RMSD values below 0.076 Å relative to their DFT-relaxed structures—almost an order of magnitude smaller than the atomic radius of hydrogen [67]. This indicates that most generated structures are very close to DFT local energy minima, reducing the need for extensive computational relaxation.

However, significant challenges remain. The CSPBench evaluation of 13 state-of-the-art algorithms demonstrates that "the performance of the current CSP algorithms is far from being satisfactory" [51]. Most algorithms struggle to identify structures with correct space groups, except for template-based approaches when applied to test structures with similar templates [51]. This highlights the continued difficulty in predicting complex crystal symmetries from first principles.

Integrated Workflows for Autonomous Discovery

End-to-End Inorganic Materials Design

The MatterGen framework demonstrates a comprehensive workflow for autonomous inorganic materials discovery:

MatterGen Workflow for Autonomous Discovery

This integrated workflow enables the generation of materials with multiple property constraints. As a proof of concept, the MatterGen team synthesized one generated structure and measured its property value to be within 20% of their target, demonstrating real-world applicability [67].

Text-Guided Exploration of Chemical Space

The Chemeleon framework introduces a novel approach to chemical space exploration by integrating textual descriptions with structural generation:

Chemeleon Text-Guided Generation Workflow

This approach supports three types of textual descriptions: composition-only (reduced composition in alphabetical order), formatted text (composition and crystal system separated by comma), and general text (diverse descriptions generated by large language models) [27]. The model demonstrated particular effectiveness in multi-component compound generation, including stable phases in the Li-P-S-Cl quaternary space relevant to solid-state batteries [27].

Organic Crystal Structure Prediction

The challenge of predicting organic crystal structures requires specialized approaches due to weaker atomic interactions and greater molecular flexibility. The SPaDe-CSP workflow addresses these challenges through machine learning-guided lattice sampling:

SPaDe-CSP Workflow for Organic Crystals

This workflow achieved an 80% success rate in tests on 20 organic crystals of varying complexity—twice that of random CSP—demonstrating how machine learning-based lattice sampling can effectively narrow the search space and increase the probability of finding experimentally observed crystal structures [31].

Research Reagent Solutions: Essential Tools for AI-Driven Discovery

Table 3: Key Research Reagents for AI-Driven Materials Discovery

Tool/Category	Specific Examples	Function	Application Context
Generative Models	MatterGen, Chemeleon, CDVAE	Generate novel crystal structures from property constraints	Inverse design of inorganic materials [67] [27]
Machine Learning Potentials	M3GNet, PFP, ANI, TeaNet	Accelerate structure relaxation with near-DFT accuracy	High-throughput screening, CSP workflows [9] [31] [51]
Benchmark Suites	CSPBench, Matbench Discovery	Standardized evaluation of algorithm performance	Method comparison, progress tracking [9] [51]
Structure Datasets	Materials Project, Alexandria, CSD	Training data for AI models	Model development, validation [67] [31]
Search Algorithms	CALYPSO, USPEX, AGOX	Global optimization of crystal structures	De novo crystal structure prediction [51]
Text-Encoding Models	Crystal CLIP, MatTPUSciBERT	Bridge textual descriptions with structural data	Text-guided materials generation [27]

Implementation Frameworks and National Initiatives

The transition from experimental algorithms to production-ready discovery platforms requires robust infrastructure and coordinated investment. The recently announced "Genesis Mission" by the U.S. government represents a comprehensive framework for scaling AI-driven discovery:

The American Science and Security Platform will integrate high-performance computing resources, AI modeling frameworks, domain-specific foundation models, and experimental tools to create an end-to-end ecosystem for autonomous discovery [68]. This infrastructure will provide "AI agents to explore design spaces, evaluate experimental outcomes, and automate workflows" [68], substantially reducing the barrier to implementation for research institutions.

Cross-sector coordination mechanisms established by the Genesis Mission include standardized partnership frameworks, clear intellectual property policies, and uniform data access standards [68]. These governance structures are essential for facilitating collaboration between academic researchers, national laboratories, and industry partners while maintaining security and maximizing public benefit.

At the organizational level, successful implementation requires fundamental workflow redesign rather than superficial automation. AI high performers are "more than three times more likely than others are to say their organizations have fundamentally redesigned individual workflows" [69]. This systematic approach to process transformation distinguishes organizations that achieve significant value from AI investments.

The integration of AI into materials discovery workflows has progressed from theoretical possibility to practical reality, though with significant limitations in certain domains. Current generative models for inorganic materials demonstrate impressive performance, with stability rates exceeding 75% and structural accuracy within 0.076 Å of DFT-optimized structures [67]. The conditioning capabilities of these models enable genuine inverse design across a broad range of property constraints, from electronic and magnetic properties to chemical composition and symmetry requirements.

However, fundamental challenges remain in the prediction of complex crystal symmetries and the accurate assessment of synthesizability beyond thermodynamic stability. The performance of current CSP algorithms is "far from being satisfactory" [51], particularly for complex multi-component systems. The disconnect between computed formation energy and real-world synthesizability represents a persistent gap that requires improved kinetic models and integration of experimental data.

The readiness of AI for autonomous discovery must be assessed domain-specifically: for inorganic materials with moderate complexity, current generative approaches offer transformative potential; for organic molecular crystals, specialized workflows like SPaDe-CSP provide significant but more limited improvements; for complex multi-component systems with specific synthesizability requirements, human-AI collaborative approaches remain essential. As benchmarking frameworks mature and infrastructure initiatives like the Genesis Mission provide production-ready platforms, the coming years will likely see accelerated adoption of integrated autonomous discovery workflows across materials research domains.

Conclusion

The journey to reliably predict synthesizable inorganic crystals is rapidly evolving from reliance on imperfect thermodynamic proxies to sophisticated, data-driven AI models. These new approaches, including deep learning networks and fine-tuned large language models, are learning the complex, implicit rules of solid-state chemistry directly from the entirety of known experimental data, achieving precision that can surpass human experts. The successful integration of synthesizability classification with precursor and synthesis-method prediction marks a pivotal shift towards closed-loop, autonomous materials discovery. For biomedical and clinical research, these advances promise to drastically accelerate the development of new functional materials, such as biocompatible coatings, drug delivery matrices, and diagnostic sensors, by ensuring that computationally designed candidates are not only high-performing but also synthetically tractable. Future progress hinges on building larger, more nuanced experimental datasets and further refining the explainability and reliability of AI models to fully bridge the gap between in-silico prediction and laboratory synthesis.

References

• 1. Machine-Learning-Assisted Prediction of Synthesizable ... [https://arxiv.org/html/2505.09161v1]
• 2. Explainable Synthesizability Prediction of Inorganic Crystal ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12051750/]
• 3. Predicting the synthesizability of crystalline inorganic ... [https://www.nature.com/articles/s41524-023-01114-4]
• 4. Autonomous platforms for data-driven organic synthesis [https://pmc.ncbi.nlm.nih.gov/articles/PMC8885738/]
• 5. Predicting synthesis recipes of inorganic crystal materials ... [https://pubs.rsc.org/en/content/articlelanding/2024/sc/d3sc03538g]
• 6. Synthesis challenges, thermodynamic stability, and growth ... [https://pubs.rsc.org/en/content/articlelanding/2025/ta/d5ta05069c]
• 7. Accurate prediction of synthesizability and precursors of 3D ... [https://www.nature.com/articles/s41467-025-61778-y]
• 8. Into the Unknown: How Computation Can Help Explore ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9585593/]
• 9. A framework to evaluate machine learning crystal stability ... [https://www.nature.com/articles/s42256-025-01055-1]
• 10. Embedding human knowledge in material screening ... [https://pubs.rsc.org/en/content/articlelanding/2025/fd/d4fd00120f]
• 11. Towards quantitative evaluation of crystal structure ... [https://www.sciencedirect.com/science/article/abs/pii/S0927025624000235]
• 12. Predicting crystal form stability under real-world conditions [https://pmc.ncbi.nlm.nih.gov/articles/PMC10632141/]
• 13. Crystal synthesizability prediction using contrastive positive ... [https://www.sciencedirect.com/science/article/abs/pii/S0010465524003886]
• 14. Solid-state synthesizability predictions using positive ... [https://pubs.rsc.org/en/content/articlehtml/2025/dd/d5dd00065c]
• 15. Explainable Synthesizability Prediction of Inorganic Crystal ... [https://www.osti.gov/biblio/2536774]
• 16. Predicting Synthesizability using Machine Learning on ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9996807/]
• 17. Accurate prediction of synthesizability and precursors of 3D ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12264133/]
• 18. Material Synthesis 2025 (MatSyn25) Dataset for 2D Materials [https://arxiv.org/abs/2510.00776]
• 19. Network analysis of synthesizable materials discovery [https://www.nature.com/articles/s41467-019-10030-5]
• 20. Predicting synthesis recipes of inorganic crystal materials ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10793203/]
• 21. 3-D Inorganic Crystal Structure Generation and Property ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7592118/]
• 22. A Synthesizability-Guided Pipeline for Materials Discovery [https://arxiv.org/html/2511.01790v1]
• 23. Predicting the synthesizability of crystalline inorganic ... [https://www.osti.gov/pages/biblio/1996869]
• 24. "Predicting Material Structures and Properties Using Deep ... [https://scholarcommons.sc.edu/etd/7481/]
• 25. Deep learning generative model for crystal structure ... [https://www.nature.com/articles/s41524-024-01443-y]
• 26. AI Finds Millions of New Stable Crystal Structures for Clean ... [https://news.sustainability-directory.com/research/ai-finds-millions-of-new-stable-crystal-structures-for-clean-energy/]
• 27. Exploration of crystal chemical space using text-guided ... [https://www.nature.com/articles/s41467-025-59636-y]
• 28. Retro-Rank-In: A Ranking-Based Approach for Inorganic ... [https://arxiv.org/html/2502.04289v1]
• 29. Chemistry-intuitive explanation of graph neural networks ... [https://www.nature.com/articles/s41467-023-38192-3]
• 30. [2407.07016] Is Large Language Model All You Need to ... [https://arxiv.org/abs/2407.07016]
• 31. Crystal structure prediction of organic molecules by machine ... [https://pubs.rsc.org/en/content/articlehtml/2025/dd/d5dd00304k]
• 32. Insightful classification of crystal structures using deep ... [https://www.nature.com/articles/s41467-018-05169-6]
• 33. Predicting melting temperature of inorganic crystals via ... [https://www.sciencedirect.com/science/article/pii/S0927025624000041]
• 34. Scaling deep learning for materials discovery [https://www.nature.com/articles/s41586-023-06735-9]
• 35. Kolmogorov–Arnold graph neural networks for molecular ... [https://www.nature.com/articles/s42256-025-01087-7]
• 36. Predicting synthesizability of crystalline materials via deep ... [https://www.nature.com/articles/s43246-021-00219-x]
• 37. A good-practice guide to solving and refining molecular ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12497095/]
• 38. SynCoTrain: a dual classifier PU-learning framework for ... [https://pubs.rsc.org/en/content/articlehtml/2025/dd/d4dd00394b]
• 39. Predicting Substrate Reactivity in Oxidative Homocoupling ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12572977/]
• 40. A Machine Learning Approach for Validating Computation ... [https://pubmed.ncbi.nlm.nih.gov/40787831/]
• 41. Advancing materials discovery through artificial intelligence [https://www.sciencedirect.com/science/article/pii/S2352940725003981]
• 42. LLMs for Explainable AI: A Comprehensive Survey [https://arxiv.org/html/2504.00125v1]
• 43. What Is AI Interpretability? | IBM [https://www.ibm.com/think/topics/interpretability]
• 44. Explainability Techniques for LLMs & AI Agents [https://testrigor.com/blog/explainability-techniques-for-llms-ai-agents/]
• 45. 11 LLMs for Explainable Artificial Intelligence: Automating ... [https://lamethods.org/book2/chapters/ch11-llmsxai/ch11-llmsxai.html]
• 46. Enabling AI to explain its predictions in plain language [https://news.mit.edu/2024/enabling-ai-explain-predictions-plain-language-1209]
• 47. Large Language Models Hallucination: A Comprehensive ... [https://arxiv.org/html/2510.06265v2]
• 48. From Illusion to Insight: A Taxonomic Survey of ... [https://www.mdpi.com/2673-2688/6/10/260]
• 49. Predicting synthesis recipes of inorganic crystal materials ... [https://pubs.rsc.org/en/content/articlehtml/2024/sc/d3sc03538g]
• 50. AtomWorld: A Benchmark for Evaluating Spatial Reasoning ... [https://arxiv.org/html/2510.04704v2]
• 51. a benchmark and critical evaluation of Crystal Structure ... [https://arxiv.org/html/2407.00733v1]
• 52. A graph theoretical approach to experimental prioritization ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11522061/]
• 53. Accelerating materials innovation with AI [https://www.innovationnewsnetwork.com/accelerating-materials-innovation-with-ai/62755/]
• 54. The AI Materials Hype Cycle Hits Its Awkward Teenage ... [https://winsomemarketing.com/ai-in-marketing/the-ai-materials-hype-cycle-hits-its-awkward-teenage-phase]
• 55. Materials Expert-Artificial Intelligence for materials discovery [https://www.nature.com/articles/s43246-025-00928-7]
• 56. WyCryst: Wyckoff inorganic crystal generator framework [https://www.sciencedirect.com/science/article/pii/S2590238524003059]
• 57. Establishing baselines for generative discovery of inorganic ... [https://pubs.rsc.org/en/content/articlehtml/2025/mh/d5mh00010f]
• 58. Impact of Data Bias on Machine Learning for Crystal ... [https://arxiv.org/html/2406.17956v1]
• 59. Synthesizability of materials stoichiometry using semi- ... [https://www.sciencedirect.com/science/article/pii/S2590238524002273]
• 60. Challenges in Non-Polymeric Crystal Structure Prediction [https://arxiv.org/html/2509.00832]
• 61. the impact of single-step retrosynthesis on synthesis planning [https://pubs.rsc.org/en/content/articlehtml/2024/dd/d3dd00252g]
• 62. Retrosynthesis prediction using an end-to-end graph ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10209957/]
• 63. RSGPT: a generative transformer model for retrosynthesis ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12314115/]
• 64. Retrosynthesis prediction with an iterative string editing ... [https://www.nature.com/articles/s41467-024-50617-1]
• 65. Retrosynthesis - TorchDrug 0.2.1 documentation [https://torchdrug.ai/docs/benchmark/retrosynthesis.html]
• 66. Crystal structure generation with autoregressive large ... [https://www.nature.com/articles/s41467-024-54639-7]
• 67. A generative model for inorganic materials design [https://www.nature.com/articles/s41586-025-08628-5]
• 68. Launching the Genesis Mission [https://www.whitehouse.gov/presidential-actions/2025/11/launching-the-genesis-mission/]
• 69. The State of AI: Global Survey 2025 [https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai]