Beyond Stability: Accuracy Metrics for Predicting Crystalline Material Synthesizability

Abstract

Accurately predicting which computationally designed crystal structures can be experimentally synthesized is a critical bottleneck in materials discovery. This article provides a comprehensive overview of the metrics and methodologies used to evaluate synthesizability predictions, moving beyond traditional thermodynamic stability measures. We explore the foundational concepts of positive-unlabeled learning, survey cutting-edge machine learning models like fine-tuned Large Language Models and graph neural networks, and detail key accuracy metrics such as true positive rate and precision. The content also addresses common challenges like data quality and model explainability, and offers a comparative analysis of different approaches. Finally, we discuss the validation of these models through experimental synthesis, providing researchers and scientists with a framework to critically assess and select the most reliable tools for accelerating the discovery of new functional materials, including those for biomedical applications.

The Synthesizability Prediction Challenge: Moving Beyond Thermodynamic Stability

Why Energy Above Hull is an Incomplete Metric for Synthesizability

The discovery of new crystalline materials is a fundamental driver of innovation across numerous scientific and technological fields, from developing better battery electrodes to creating novel superconductors. A critical step in this process is determining whether a computationally predicted material can be successfully synthesized in a laboratory. For years, the energy above hull (Eₕᵤₗₗ) has served as a primary thermodynamic proxy for assessing synthesizability. This metric represents a material's energy relative to the most stable phases in its composition space, with values near zero typically interpreted as indicating stability and thus potential synthesizability. However, a growing body of evidence demonstrates that Eₕᵤₗₗ alone provides an incomplete picture of synthesizability, leading to both false positives (materials predicted to be synthesizable that are not) and false negatives (overlooking metastable materials that can be synthesized). This limitation has prompted the development of sophisticated machine learning approaches that capture the complex, multi-faceted nature of materials synthesis beyond simple thermodynamic considerations.

Fundamental Limitations of Energy Above Hull

Theoretical Shortcomings of a Pure Thermodynamic Metric

The energy above hull metric suffers from several fundamental limitations that restrict its utility as a comprehensive synthesizability indicator. First, Eₕᵤₗₗ is fundamentally a thermodynamic metric calculated at zero Kelvin, which ignores crucial kinetic factors that govern real-world synthesis outcomes. While materials with low Eₕᵤₗₗ values are thermodynamically favored, their synthesis may be impeded by high activation energy barriers that prevent formation from available precursors. Conversely, many metastable materials with positive Eₕᵤₗₗ values can be synthesized through kinetic stabilization, where they remain trapped in local energy minima despite not being the global ground state [1].

Second, Eₕᵤₗₗ fails to account for technological and experimental constraints that significantly impact synthesis success. The ability to synthesize a material often depends on available equipment, precursor availability, specific reaction conditions, and the current state of synthetic methodology. For instance, novel high-entropy alloys with significant potential for catalysis applications were recently synthesized using the Carbothermal Shock method, achieving homogeneous components and uniform structures that were inaccessible through conventional synthesis techniques [1]. Similarly, some materials can only be synthesized under extreme conditions, such as high pressure, despite having favorable formation energies under standard conditions [1].

Third, vibrational stability represents another crucial factor overlooked by Eₕᵤₗₗ analysis. Materials can exhibit favorable Eₕᵤₗₗ values yet be vibrationally unstable, as indicated by imaginary phonon modes in their vibrational spectra. For example, LiZnPS₄ (mp-11175) with Eₕᵤₗₗ = 0 meV, SiC (mp-11713) with Eₕᵤₗₗ = 3 meV, and Ca₃PN (mp-11824) with Eₕᵤₗₗ = 0 meV all demonstrate vibrational instability despite their apparently favorable thermodynamic profiles [2].

Practical Limitations in Materials Discovery

Beyond theoretical limitations, Eₕᵤₗₗ faces practical challenges in guiding materials discovery. The metric cannot differentiate between polymorphs of the same composition, despite their potentially vastly different synthetic accessibility. Additionally, Eₕᵤₗₗ provides no guidance on appropriate synthesis routes, precursors, or reaction conditions—essential information for experimentalists. The Materials Project lists 21 SiO₂ structures within 0.01 eV of the convex hull, yet the second most common phase, cristobalite (β-quartz), is not among these, highlighting the disconnect between thermodynamic stability and actual synthetic prevalence [3].

Traditional heuristic approaches like the Pauling Rules or charge-balancing criteria have also proven insufficient for synthesizability prediction. More than half of the experimental materials in the Materials Project database do not meet these established criteria, further underscoring the need for more sophisticated assessment methods [1].

Emerging Machine Learning Approaches

Machine learning models have emerged as powerful alternatives to Eₕᵤₗₗ-based synthesizability assessment, capable of integrating diverse chemical, structural, and experimental factors that influence synthesis outcomes.

Key Methodological Frameworks

Positive-Unlabeled (PU) Learning represents a particularly significant advancement, as it directly addresses the fundamental data challenge in synthesizability prediction: the absence of confirmed negative examples. Since failed synthesis attempts are rarely published, ML models cannot access reliable "unsynthesizable" examples for training. PU learning frameworks treat all non-synthesized materials as "unlabeled" rather than definitively unsynthesizable, then iteratively identify the most likely negative examples from this pool. This approach has been successfully implemented in various architectures, including graph neural networks and large language models [1] [4].

Co-training frameworks like SynCoTrain leverage multiple complementary models to reduce individual model bias and enhance generalizability. SynCoTrain employs two distinct graph convolutional neural networks—SchNet and ALIGNN—that iteratively exchange predictions. SchNet uses continuous convolution filters suitable for encoding atomic structures (a "physicist's perspective"), while ALIGNN directly encodes atomic bonds and bond angles (a "chemist's perspective"). This collaborative approach improves reliability for out-of-distribution predictions, which is crucial for identifying truly novel materials [1].

Large Language Models (LLMs) have recently demonstrated remarkable capabilities in synthesizability prediction. The Crystal Synthesis LLM (CSLLM) framework utilizes specialized language models fine-tuned on text representations of crystal structures to predict synthesizability, synthetic methods, and suitable precursors. By representing crystal structures as human-readable text descriptions, these models can leverage patterns learned from vast chemical literature corpora [5] [4].

Comparative Performance of ML Models

Table 1: Performance Comparison of Synthesizability Prediction Methods

Method	Approach	Key Features	Reported Accuracy/Performance
Energy Above Hull	Thermodynamic	Distance from convex hull	Limited by ignoring kinetic and technological factors
Charge-Balancing	Heuristic	Net neutral ionic charge	Only 37% of synthesized materials are charge-balanced [6]
SynCoTrain [1]	Dual-classifier PU-learning	Co-training with SchNet & ALIGNN	High recall on oxide crystals
SynthNN [6]	Deep learning (composition-based)	atom2vec composition embeddings	7× higher precision than formation energy [6]
CSLLM [5]	Fine-tuned LLM	Material string representation	98.6% accuracy [5]
PU-GPT-embedding [4]	LLM embeddings + PU-learning	Text-embedding-3-large representations	Outperforms graph-based methods

Table 2: Experimental Validation of ML-Guided Discovery Pipelines

Study	Approach	Candidates Screened	Experimentally Validated	Success Rate
Prein et al. [3]	Composition + structure rank-average ensemble	4.4 million structures	7 of 16 targets synthesized	44%
CSLLM Framework [5]	Multi-task LLM prediction	105,321 theoretical structures	45,632 identified as synthesizable	High-throughput screening

Experimental Protocols and Methodologies

Data Curation and Representation

Each ML approach employs specialized data curation strategies to address the unique challenges of synthesizability prediction. The PU learning framework typically uses confirmed synthesized materials from databases like the Inorganic Crystal Structure Database (ICSD) as positive examples, while treating hypothetical materials from computational databases (Materials Project, OQMD, JARVIS) as unlabeled data [5]. For structure-based models, crystal graphs represent atoms as nodes and bonds as edges, capturing structural relationships directly [1]. Composition-based models like SynthNN utilize learned atom embeddings (atom2vec) that optimize feature representation alongside other model parameters [6].

The CSLLM framework introduces a novel "material string" representation that efficiently encodes crystal structures as text by including space group information, lattice parameters, and Wyckoff positions while eliminating redundant atomic coordinates [5]. This representation enables the application of LLMs to crystal structure analysis. Similarly, image-based representations color-code chemical attributes into 3D pixel-wise images, allowing convolutional neural networks to learn hidden synthesizability features from visual patterns [7].

Model Architectures and Training

SynCoTrain implements a semi-supervised co-training framework where two GCNNs (SchNet and ALIGNN) iteratively refine predictions on unlabeled data. SchNet employs continuous-filter convolutional layers that model atomic interactions through learned energy functions, while ALIGNN explicitly represents both bond and angle information in its graph structure. The models alternate training epochs and exchange high-confidence predictions to expand each other's training sets, progressively improving decision boundaries [1].

LLM-based approaches like CSLLM fine-tune foundation models (GPT-4o-mini) on text descriptions of crystal structures generated by tools like Robocrystallographer. The fine-tuning process adapts the models' general language capabilities to the specific domain of crystal structure analysis, enabling them to recognize synthesizability patterns from structural descriptions [4]. For enhanced performance, LLM-generated embeddings can be used as input to dedicated PU-classifier networks rather than using the LLMs as direct classifiers.

Ensemble methods combine compositional and structural signals through separate encoders—typically a transformer for composition and a graph neural network for structure—with rank-average fusion of their predictions. This approach acknowledges that synthesizability depends on both elemental chemistry (precursor availability, redox constraints) and structural features (local coordination, motif stability) [3].

Visualization of Methodologies and Relationships

Synthesizability Factors and ML Approaches Diagram

ML Workflow for Synthesizability Prediction

Table 3: Key Computational Tools and Databases for Synthesizability Research

Resource	Type	Primary Function	Relevance to Synthesizability
Materials Project [1] [8]	Database	DFT-calculated material properties	Source of Eₕᵤₗₗ values and crystal structures for training
ICSD [6] [5]	Database	Experimentally confirmed structures	Source of positive examples for ML training
Robocrystallographer [4]	Software Tool	Generates text descriptions of crystals	Creates LLM-readable input from CIF files
ALIGNN [1]	ML Model	Graph neural network with angle information	Captures bond angles in addition to atomic connections
SchNet [1]	ML Model	Continuous-filter convolutional network	Models quantum interactions in atomic systems
PU-CGCNN [4] [9]	ML Framework	Positive-unlabeled crystal graph convolutional net	Addresses lack of negative examples in training data
CSLLM [5]	ML Framework	Specialized large language models	Predicts synthesizability, methods, and precursors

The evidence clearly demonstrates that energy above hull provides an incomplete metric for synthesizability prediction due to its fundamental limitation as a pure thermodynamic measure. While valuable for assessing thermodynamic stability, Eₕᵤₗₗ fails to capture kinetic barriers, technological constraints, vibrational stability, and polymorph-specific synthetic accessibility that ultimately determine whether a material can be successfully synthesized. Machine learning approaches—including PU learning, co-training frameworks, and large language models—offer powerful alternatives that integrate diverse data sources and capture complex patterns beyond thermodynamic considerations. These methods have demonstrated superior performance in both computational benchmarks and experimental validation, successfully guiding the synthesis of novel materials that would have been overlooked by Eₕᵤₗₗ-based screening alone. The future of synthesizability prediction lies in combining these data-driven approaches with physical insights, creating hybrid models that leverage both computational efficiency and scientific understanding to accelerate functional materials discovery.

A silent revolution is underway in materials science. For decades, the discovery of new inorganic crystalline materials has been hampered by a fundamental bottleneck: determining which computationally designed compounds can be successfully synthesized in the laboratory. While high-throughput computational methods now generate millions of promising candidate materials with desirable properties, the vast majority prove impossible to synthesize through known methods [5]. This challenge stems from a fundamental gap in our data ecosystems—the scarcity of reliably labeled 'non-synthesizable' examples, without which machine learning models cannot effectively learn the complex constraints governing successful synthesis.

The prediction of material synthesizability represents a critical bridge between theoretical materials design and experimental realization [10]. Traditional approaches have relied on proxy metrics like thermodynamic stability (energy above the convex hull) or charge-balancing principles, but these have proven insufficient [6] [11]. Materials with favorable formation energies often remain unsynthesized, while numerous metastable structures are routinely synthesized despite less favorable thermodynamics [5]. The development of accurate synthesizability predictors therefore requires moving beyond these proxies to learn directly from the complete distribution of synthesized materials—and crucially, from their negative counterparts.

This comparison guide examines the core methodologies emerging to address the fundamental challenge of defining and sourcing 'non-synthesizable' data. We objectively compare the performance, experimental protocols, and underlying assumptions of three dominant approaches: Positive-Unlabeled (PU) Learning, Human-Curated Datasets, and Large Language Models (LLMs). By synthesizing quantitative comparisons and detailed methodological analyses, we provide researchers with a framework for evaluating and selecting appropriate strategies for synthesizability prediction in their own materials discovery workflows.

Methodological Comparison: Defining the Undefined

Positive-Unlabeled (PU) Learning Approaches

Core Principle: PU learning frameworks treat the lack of synthesis evidence not as definitive negative labels, but as "unlabeled" examples that may include both synthesizable and non-synthesizable materials. These methods probabilistically weight unlabeled examples during training according to their likelihood of being synthesizable [6] [12].

Experimental Protocol: The standard implementation involves:

• Positive Set Curation: Collecting experimentally synthesized materials from authoritative databases like the Inorganic Crystal Structure Database (ICSD) [6] [5] [12].
• Unlabeled Set Generation: Creating a candidate set of theoretically possible but experimentally unreported materials from computational databases (Materials Project, OQMD, AFLOW) [5] [10].
• Model Training: Implementing specialized algorithms that treat synthesized materials as positive examples and all others as unlabeled, with class-weighted loss functions that account for the unknown label distribution [6] [12].

Representative Models: SynthNN (deep learning synthesizability model) [6], CLscore (crystal-likeness score) [5], and various semi-supervised implementations [12].

Table 1: Performance Metrics of PU Learning Models

Model	Accuracy/Precision	Recall/True Positive Rate	Key Advantages	Limitations
SynthNN [6]	7× higher precision than DFT formation energy	Not specified	Learns chemical principles without prior knowledge; 5 orders of magnitude faster than human experts	Cannot definitively label materials as unsynthesizable
CLscore [5]	87.9% accuracy (3D crystals)	Not specified	Effective for screening large theoretical databases	Limited by quality of underlying computational structures
Semi-Supervised [12]	83.6% estimated precision	83.4%	Enables continuous synthesizability phase mapping across compositional spaces	Performance varies across material systems

Human-Curated Literature Datasets

Core Principle: This approach involves manual extraction of synthesis information from scientific literature, including explicit records of both successful and failed synthesis attempts. This provides explicitly labeled negative examples rather than relying on algorithmic inference [11].

Experimental Protocol: The meticulous curation process involves:

• Database Querying: Identifying candidate materials from computational databases (e.g., ternary oxides from Materials Project) with associated literature references [11].
• Manual Literature Review: Systematically examining primary sources (journal articles, ICSD records) to determine synthesis outcomes through explicit statements or experimental details [11].
• Data Extraction & Labeling: Categorizing materials as "solid-state synthesized," "non-solid-state synthesized," or "undetermined" based on documented evidence, with additional metadata collection on reaction conditions [11].

Implementation Example: A recent study manually curated 4,103 ternary oxides, identifying 3,017 as solid-state synthesized, 595 as non-solid-state synthesized, and 491 as undetermined due to insufficient evidence [11].

Table 2: Human-Curated Dataset Applications

Application	Dataset Size	Key Findings	Validation Method
Solid-State Synthesizability Prediction [11]	4,103 ternary oxides	Identified 156 outliers in text-mined datasets; predicted 134/4312 hypothetical compositions as synthesizable	100 randomly chosen entries validated by independent researcher
Synthesis Condition Analysis [11]	3,017 solid-state synthesized entries	Enabled correlation of heating temperatures with precursor melting points	Cross-referenced with established materials databases
Text-Mining Validation [11]	4,800 text-mined entries	Only 15% of outliers correctly extracted in automated pipelines	Manual verification of synthesis descriptions

Large Language Models (LLMs) for Synthesizability

Core Principle: Leveraging pre-trained LLMs fine-tuned on comprehensive datasets of both synthesizable and non-synthesizable crystal structures, using specialized text representations of material information [5].

Experimental Protocol: The CSLLM framework implements:

• Balanced Dataset Construction: 70,120 synthesizable structures from ICSD paired with 80,000 non-synthesizable structures identified via PU learning pre-screening [5].
• Text Representation: Development of "material string" format that condenses essential crystal information (space group, lattice parameters, atomic coordinates) [5].
• Specialized Model Fine-tuning: Training three separate LLMs for synthesizability prediction, synthetic method classification, and precursor identification [5].

Performance Highlights: The Synthesizability LLM achieves 98.6% accuracy, significantly outperforming traditional thermodynamic (74.1%) and kinetic (82.2%) stability metrics [5].

Performance Benchmarking Across Material Systems

Quantitative Comparison of Prediction Accuracy

Different methodological approaches show varying performance characteristics across material systems and evaluation metrics. The table below synthesizes direct comparisons where available and contextualizes results across studies.

Table 3: Cross-Method Performance Benchmarking

Method	Material System	Accuracy	Precision	Recall/TPR	Key Innovation
LLM (CSLLM) [5]	3D crystals (70,120 structures)	98.6%	Not specified	Not specified	Material string representation; multi-task learning
PU Learning [12]	Inorganic compositions	Not specified	83.6% (estimated)	83.4%	Continuous synthesizability phase mapping
Synthesizability Score [10]	Ternary crystals	82.6%	82.6%	80.6%	Fourier-transformed crystal properties (FTCP)
Human Expert [6]	Various inorganic materials	Not specified	1.5× lower than SynthNN	Not specified	Domain expertise and literature knowledge
Charge-Balancing [6]	Known synthesized materials	37% of known materials charge-balanced	Not specified	Not specified	Simple heuristic based on oxidation states

Experimental Validation and Real-World Discovery

Beyond quantitative metrics, the most significant validation of synthesizability prediction methods comes from experimental confirmation of novel materials discoveries.

PU Learning Guided Discovery: In one implementation, a semi-supervised learning model successfully guided experimental exploration of quaternary oxide compositional space (CuO, Fe₂O₃, V₂O₅), resulting in the discovery of a new phase, Cu₄FeV₃O₁₃ [12]. This demonstrates the practical utility of synthesizability predictions in directing resource-intensive experimental efforts toward promising compositional regions.

Temporal Validation: Another approach trained a synthesizability score model exclusively on materials reported before 2015, then tested on compounds added to databases after 2019. The model achieved an 88.60% true positive rate, coupled with 9.81% precision, indicating that newly added materials remained unexplored and had high synthesis potential [10]. This temporal validation approach provides strong evidence for the predictive capability of these models beyond simple reproduction of known data.

Research Reagent Solutions: Essential Materials & Tools

Table 4: Key Experimental and Computational Resources

Resource	Type	Function	Example Sources
ICSD [6] [5] [11]	Database	Authoritative source of experimentally synthesized crystalline structures	FIZ Karlsruhe
Materials Project [5] [11] [10]	Database	DFT-calculated structures and properties for synthesized and hypothetical materials	LBNL materialsproject.org
OQMD/AFLOW [5] [10]	Database	Additional sources of theoretical structures for negative example generation	University of Chicago, Duke University
PU Learning Algorithms [6] [12]	Software Framework	Handles incomplete negative labeling through semi-supervised approaches	Custom implementations in Python
Text-Mining Pipelines [11]	Data Processing	Automated extraction of synthesis information from literature	Natural language processing tools
LLM Fine-tuning [5]	Computational Method	Adapts general language models to crystal structure prediction	Transformer architectures (LLaMA, etc.)

The critical challenge of defining and sourcing 'non-synthesizable' data has spawned diverse methodological approaches, each with distinct strengths and limitations. PU learning frameworks offer scalability and effectiveness with large-scale computational databases but cannot definitively label materials as unsynthesizable. Human-curated datasets provide high-quality, explicit negative examples but face significant scalability constraints. LLM-based approaches demonstrate remarkable accuracy when trained on balanced, pre-screened datasets but require specialized text representations and substantial computational resources.

For researchers navigating this landscape, selection criteria should include: the scale of the target material space, availability of domain expertise for manual curation, computational resources, and the requirement for synthesis route prediction beyond binary synthesizability classification. As these methodologies continue to evolve, the integration of their complementary strengths—perhaps through ensemble approaches or hybrid human-AI curation systems—promises to further accelerate the discovery of synthesizable functional materials.

The progression from proxy metrics to data-driven predictors represents a paradigm shift in materials discovery, directly addressing the central problem of distinguishing viable candidates from the vast chemical space of non-synthesizable possibilities. This capability will prove increasingly vital as computational materials design continues to outpace experimental validation, ensuring that theoretical promise translates to practical realization.

Predicting which theoretically designed materials can be successfully synthesized in the laboratory remains a grand challenge in materials science. Traditional proxies for synthesizability, such as thermodynamic and kinetic stability, often fail to capture the complex realities of experimental synthesis. Positive-Unlabeled (PU) learning has emerged as a powerful machine learning framework to address this, enabling accurate synthesizability predictions by learning only from known synthesized ("positive") materials and a large set of "unlabeled" theoretical candidates. This guide provides a comprehensive comparison of PU learning methodologies, performance metrics, and experimental protocols specifically for crystalline material synthesizability prediction, examining how different algorithmic approaches achieve state-of-the-art accuracy where traditional methods fall short.

The discovery of new functional materials is crucial for advancing technologies in energy storage, electronics, and sustainability. While computational methods can rapidly screen thousands of theoretical material designs, experimental validation remains a critical bottleneck. This challenge is compounded by the fundamental asymmetry in materials data: we have extensive records of successfully synthesized materials but scarce data on failed synthesis attempts. Materials databases contain well-documented positive examples, but definitive negative examples are rarely reported in scientific literature [11] [6].

Traditional synthesizability screening relies heavily on thermodynamic stability metrics, particularly energy above the convex hull (Ehull), which measures a material's stability relative to its potential decomposition products. However, this approach has significant limitations. Studies show that a non-negligible number of hypothetical materials with low Ehull have not been synthesized, while many metastable materials with higher E_hull have been successfully synthesized [11]. Kinetic barriers, entropic contributions, and specific synthesis conditions further complicate the relationship between thermodynamic stability and actual synthesizability.

PU learning reframes this challenge as a weakly supervised binary classification problem where the goal is to learn a binary classifier from only positive and unlabeled data, without access to confirmed negative examples [13]. This approach aligns perfectly with the realities of materials data, where we have confirmed positive examples (known synthesized materials) and numerous unlabeled candidates (theoretical materials with unknown synthesizability).

PU Learning Fundamentals

Problem Formulation and Key Assumptions

In formal terms, PU learning aims to learn a binary classifier (f: \mathcal{X} \rightarrow \mathbb{R}) from a positive training set (DP = {(\boldsymbol{x}i, +1)}{i=1}^{nP}) and an unlabeled training set (DU = {\boldsymbol{x}i}{i=nP+1}^{nP+nU}), where (\mathcal{X} \subseteq \mathbb{R}^d) is the feature space [13]. The key challenge is that the unlabeled set contains both positive and negative instances, but without distinguishing labels.

Two primary data generation assumptions underlie different PU learning approaches:

• One-Sample (OS) Setting: Positive and unlabeled training sets are generated sequentially from the marginal distribution, with positive labels observed with constant probability [13].
• Two-Sample (TS) Setting: Positive and unlabeled training sets are generated independently from their respective distributions [13].

These settings have important practical implications. The OS setting more closely resembles real-world materials data collection, while many algorithms are designed for the TS setting. Recent research has identified that failing to account for this distinction can lead to unfair performance comparisons and suboptimal results [13].

Algorithmic Families in PU Learning

PU learning algorithms have evolved into three main families, each with distinct approaches to handling the missing negative information:

Table 1: PU Learning Algorithm Families

Algorithm Family	Core Mechanism	Key Advantages	Materials Science Applications
Cost-Sensitive	Assigns different weights to positive and unlabeled data to approximate classification risk [13]	Theoretical risk consistency; No explicit negative selection needed	General synthesizability prediction [6]
Sample-Selection	Identifies high-confidence negative examples from unlabeled data for supervised learning [13]	Leverages existing supervised algorithms; Interpretable negative selection	MXene synthesizability prediction [14]
Biased Learning	Models the biased generation process of positive data with correction approaches [13]	Accounts for selection bias in positive labeling	Solid-state synthesizability prediction [11]

PU Learning Approaches for Materials Synthesizability

Neighborhood-Based Methods with Decision Trees

The Neighborhood-Based Positive Unlabeled Learning Using Decision Tree (NPULUD) method represents a recent advance combining nearest-neighbor analysis with decision tree classification. This approach uses the k-nearest neighbors algorithm for PU strategy and employs decision trees with entropy measures for classification [15]. Entropy serves as a crucial measure for assessing uncertainty in the training dataset during decision tree construction.

In comprehensive evaluations across 24 real-world datasets, NPULUD achieved an average accuracy of 87.24%, significantly outperforming traditional supervised learning approaches (83.99%) and demonstrating a 7.74% average improvement over state-of-the-art peers [15]. The method also excelled in precision (0.8572), recall (0.8724), and F-measure (0.8625) metrics, with statistical significance confirmed by Wilcoxon tests (p-value = 0.0004693) [15].

Transductive Bagging Approaches

Transductive bagging represents another powerful approach for material synthesizability prediction, particularly for 2D materials like MXenes. This method adapts a framework where some unlabeled examples are randomly labeled as "negative," then a classifier (typically decision trees) is trained to distinguish positive and negative examples [14]. Through bootstrapping—creating random subsets of the original data with replacement—the process repeats with different negative example sets until the classifier excels at recognizing positive instances.

In practice, this approach enabled the discovery of 18 new potentially synthesizable MXenes by learning complex patterns in atomic arrangements and electron distributions that go beyond simple thermodynamic considerations [14]. The model achieved a remarkable true positive rate of 0.91 across the entire Materials Project database, correctly identifying already-synthesized materials 91% of the time [14].

Large Language Models for Synthesizability Prediction

The Crystal Synthesis Large Language Models (CSLLM) framework represents a cutting-edge approach leveraging specialized LLMs fine-tuned for materials science. This framework utilizes three specialized models for predicting synthesizability, synthetic methods, and suitable precursors respectively [16]. By representing crystal structures as text using a novel "material string" representation, CSLLM achieves unprecedented 98.6% accuracy in synthesizability prediction, significantly outperforming traditional stability-based methods (E_hull ≥0.1 eV/atom: 74.1%; phonon spectrum ≥ -0.1 THz: 82.2%) [16].

The framework was trained on a balanced dataset of 70,120 synthesizable crystal structures from the Inorganic Crystal Structure Database (ICSD) and 80,000 non-synthesizable structures identified through PU learning screening of over 1.4 million theoretical structures [16]. This demonstrates how PU learning can create high-quality negative examples for training even more accurate supervised models.

Performance Comparison and Benchmarking

Quantitative Performance Metrics

Table 2: Performance Comparison of Synthesizability Prediction Methods

Method	Accuracy	Precision	Recall	F1-Score	Materials Scope	Reference
NPULUD	87.24%	0.8572	0.8724	0.8625	General (24 datasets)	[15]
CSLLM	98.6%	N/A	N/A	N/A	3D Crystals	[16]
PU Learning (Jang et al.)	87.9%	N/A	N/A	N/A	3D Crystals	[16]
Teacher-Student Network	92.9%	N/A	N/A	N/A	3D Crystals	[16]
SynthNN	7× higher precision than E_hull	N/A	N/A	N/A	Inorganic Compositions	[6]
Traditional E_hull	~50% of synthesized materials captured	N/A	N/A	N/A	General	[6]
Charge-Balancing	37% of synthesized materials captured	N/A	N/A	N/A	Inorganic Compositions	[6]

Benchmarking Challenges and Considerations

Recent research has highlighted critical challenges in fairly evaluating PU learning algorithms. Many algorithms rely on validation sets containing negative data—an unrealistic requirement in true PU settings where no confirmed negative examples exist [13]. This creates an evaluation paradox that contradicts the original motivation of PU learning.

The 2025 benchmark study by Wang et al. also identified the "internal label shift" problem, where differences between the one-sample and two-sample settings significantly impact algorithm performance [13]. Their findings revealed that no single PU learning algorithm outperforms all others on every dataset or metric, and early simple methods often achieve strong classification performance [13]. This underscores the importance of context-specific algorithm selection rather than seeking a universal best solution.

Experimental Protocols and Implementation

Data Collection and Curation

The foundation of effective PU learning for materials synthesizability lies in rigorous data curation. The protocol for solid-state synthesizability prediction involves:

• Extraction of Known Materials: 4,103 ternary oxides were manually curated from the Materials Project database with Inorganic Crystal Structure Database (ICSD) IDs, excluding non-metal elements and silicon [11].

• Literature Validation: Each ternary oxide was verified through exhaustive literature review examining ICSD records, Web of Science (first 50 results sorted by oldest to newest), and Google Scholar (top 20 relevant results) [11].

• Labeling Protocol: Materials were categorized as "solid-state synthesized" (3,017 entries), "non-solid-state synthesized" (595 entries), or "undetermined" (491 entries) based on explicit synthesis evidence [11].

• Feature Engineering: Calculation of thermodynamic, structural, and electronic properties using tools like Matminer for featurization [14].

Model Training and Validation

Effective PU learning implementation requires careful attention to model selection and validation strategies:

• Two-Step Validation: For solid-state synthesizability prediction, 100 randomly chosen entries were validated for solid-state synthesized materials, while all non-solid-state entries were checked [11].

• PU-Specific Model Selection: Employ validation criteria that use only positive and unlabeled data, avoiding the unrealistic requirement of negative examples for validation [13].

• Class Prior Estimation: Accurately estimate the proportion of positive instances in the unlabeled data, as this significantly impacts algorithm performance [13].

• Cross-Family Algorithm Testing: Evaluate both one-sample and two-sample algorithms with appropriate calibration to ensure fair comparisons [13].

Research Reagent Solutions

Table 3: Essential Computational Tools for PU Learning in Materials Science

Tool/Resource	Function	Application Example	Access
Materials Project API	Provides computational data for known and theoretical materials	Feature calculation for synthesizability prediction	materialsproject.org
Matminer	Materials feature extraction and visualization	Featurizing compositions and structures for ML	Python library
pumml Python Package	PU learning implementation specifically for materials science	Predicting synthesizability of new compounds	GitHub Repository
ICSD Database	Source of confirmed synthesized materials	Positive examples for PU training	Commercial license
Text-Mined Synthesis Datasets	Literature-derived synthesis information	Training data for method and precursor prediction	Kononova et al. 2019 [11]

PU learning has fundamentally transformed the paradigm of synthesizability prediction in materials science by directly addressing the fundamental asymmetry in materials data. Through various implementations—from neighborhood-based methods with decision trees to advanced large language models—PU learning consistently demonstrates superior performance compared to traditional stability-based approaches.

The key insights from comparative analysis reveal that while PU learning methods generally outperform traditional approaches, algorithm selection must be context-dependent. Simple early methods often remain competitive with newer approaches, and practical considerations like validation strategy and data curation quality significantly impact real-world performance. Future developments will likely focus on standardized benchmarking, improved model selection criteria without negative examples, and integration with autonomous experimentation systems.

For materials researchers implementing PU learning, success depends on rigorous data curation, appropriate algorithm selection for specific materials classes, and careful attention to validation protocols that reflect real-world constraints. As the field matures, PU learning promises to significantly accelerate materials discovery by providing reliable synthesizability assessments that bridge the gap between computational design and experimental realization.

The discovery of new functional materials is a cornerstone of technological advancement, yet the experimental realization of computationally predicted crystals remains a major bottleneck. This challenge has spurred the development of machine learning models to predict crystalline material synthesizability—whether a hypothetical material can be experimentally synthesized. However, a fundamental problem persists in evaluating these models: the absence of definitive negative examples. While databases contain confirmed synthesizable (positive) materials, truly unsynthesizable materials are rarely documented, creating an evaluation paradigm known as Positive-Unlabeled (PU) learning. This framework severely constrains the standard metrics available for model assessment, making True Positive Rate (TPR or recall) often the only reliable metric, while precision and false positive rates must be estimated with inherent uncertainty. This article examines the key performance indicators used across different synthesizability prediction approaches, compares their reported results, and discusses the critical limitations of current evaluation methodologies that researchers must navigate.

Performance Metrics Comparison of Synthesizability Prediction Models

The field has seen rapid evolution from traditional thermodynamic approaches to specialized machine learning models. The table below synthesizes quantitative performance data across major model architectures, highlighting their reported capabilities under the constraints of PU evaluation.

Table 1: Performance comparison of crystalline material synthesizability prediction models

Model / Approach	Reported True Positive Rate (TPR/Recall)	Estimated Precision	Key Evaluation Notes	Source
CSLLM (LLM-based)	Not explicitly stated	Not explicitly stated	Achieves 98.6% overall accuracy on a balanced dataset	[16]
SynthNN (Composition-based)	Not explicitly stated	7× higher than DFT formation energy	Outperformed 20 human experts in discovery precision	[6]
Teacher-Student DNN (TSDNN)	92.9%	Not explicitly stated	Improved baseline PU learning TPR from 87.9%; uses 1/49 model parameters	[17]
Perovskite GNN (Transfer Learning)	95.7%	Not explicitly stated	Domain-specific transfer learning significantly outperformed general model (74.0% TPR)	[18]
PU-CGCNN (Structure-based)	~87%	Requires α-estimation	Early structure-based PU learning benchmark	[4] [18]
Energy Above Hull (Stability)	Not applicable	Not applicable	Captures only ~50% of synthesized materials; poor synthesizability proxy	[6]
Charge Balancing	Not applicable	Not applicable	Only 37% of known synthesized materials are charge-balanced	[6]

Key Performance Metric Insights

• True Positive Rate (TPR) as Primary Metric: Due to the PU learning constraint, TPR is the most commonly reported and reliable metric, representing a model's ability to correctly identify known synthesizable materials held out from training. The progression from ~87% TPR in earlier PU-learning models to >92% in advanced architectures like TSDNN and >95% in domain-specific implementations demonstrates significant methodological improvement [17] [18].

• The Precision Estimation Challenge: Estimating precision requires α-estimation techniques, as true negative examples are unavailable [4]. While SynthNN reports 7× higher precision than DFT-based formation energy screening, such comparisons are necessarily approximate [6]. The high accuracy (98.6%) reported by CSLLM comes from evaluation on a balanced dataset with presumed negative examples, a methodology not applicable to real-world discovery scenarios [16].

• Domain-Specific Enhancements: Performance varies significantly across material classes. General models achieve ~74% TPR for perovskites, while domain-specialized versions reach 95.7%, highlighting the importance of chemical domain expertise in model architecture [18].

Experimental Protocols and Methodologies

Standard PU Learning Framework

Most synthesizability prediction models follow a consistent experimental protocol based on the PU learning paradigm. The workflow begins with data preparation from crystallographic databases like the Materials Project (MP) and Inorganic Crystal Structure Database (ICSD). Materials with ICSD IDs are typically treated as positive (synthesized) examples, while those without ICSD IDs are considered unlabeled [11] [18]. The core learning process involves iterative training where models learn to distinguish positive examples from randomly sampled unlabeled data, with multiple iterations refining the decision boundary [18]. Performance evaluation primarily relies on hold-out testing, where a subset of known positive materials (e.g., 10%) is reserved for calculating the True Positive Rate [18].

Diagram: Standard PU Learning Workflow for Synthesizability Prediction

Model-Specific Methodological Variations

• Large Language Models (CSLLM): The Crystal Synthesis LLM framework employs a balanced dataset of 70,120 synthesizable structures from ICSD and 80,000 non-synthesizable structures screened via a pre-trained PU learning model. It uses a specialized "material string" text representation of crystal structures for fine-tuning, integrating lattice parameters, composition, atomic coordinates, and symmetry information [16].

• Teacher-Student Dual Neural Network (TSDNN): This approach uses a dual-network architecture where a teacher model provides pseudo-labels for unlabeled data, which a student model then learns from. This semi-supervised approach effectively exploits large amounts of unlabeled data, achieving high TPR with significantly reduced model parameters compared to earlier implementations [17].

• Domain-Specific Transfer Learning: For perovskite prediction, researchers first pre-train a model on the general MP database, then fix weights in the encoding and first graphical convolution layers while retraining the remaining layers on a specialized perovskite dataset. This transfer learning approach increases TPR from 74.0% to 95.7% for perovskite materials [18].

• LLM-Embedding Hybrids: Some recent approaches use GPT embeddings (text-embedding-3-large) to convert crystal structure descriptions into 3072-dimensional vector representations, then apply traditional PU-classifier neural networks. This approach reportedly outperforms both fine-tuned LLMs and graph-based representations [4].

The Scientist's Toolkit: Essential Research Reagents

Table 2: Key computational tools and data resources for synthesizability prediction research

Resource	Type	Primary Function	Application in Synthesizability
Materials Project (MP)	Database	Repository of computed material properties	Source of hypothetical/unlabeled structures; formation energy data	[11] [18]
Inorganic Crystal Structure Database (ICSD)	Database	Experimentally confirmed crystal structures	Source of positive (synthesizable) examples	[11] [18]
PyMatgen	Software	Python materials analysis library	Structure manipulation, feature extraction, compatibility with MP API	[11] [18]
Robocrystallographer	Software	Text description generator for crystals	Converts CIF files to text prompts for LLM-based approaches	[4]
Crystal Graph Convolutional Neural Network (CGCNN)	Model Architecture	Graph representation of crystal structures	Base architecture for many structure-based prediction models	[17] [4] [18]
OpenAI GPT Models	Foundation Models	Large language models	Fine-tuned for synthesizability classification or embedding generation	[4]
Positive-Unlabeled Learning Algorithms	Methodology	Semi-supervised classification	Core learning framework for handling lack of negative examples	[11] [17] [18]

Critical Analysis of PU Evaluation Limitations

The Positive-Unlabeled learning framework fundamentally limits the assessment of synthesizability prediction models, creating several critical challenges for the field.

The True Negative Data Deficiency

The most significant limitation is the absence of confirmed unsynthesizable materials. As noted in the human-curated study of ternary oxides, scientific literature rarely reports failed synthesis attempts [11]. This absence means:

• Precision Estimation Relies on α-estimation: Techniques like α-estimation must be used to approximate precision and false positive rates, introducing uncertainty [4].
• No Direct False Positive Measurement: Models cannot be directly evaluated on their ability to reject truly unsynthesizable compounds, only on their recall of known synthesizable ones.
• Artificial Negative Sampling: Some approaches, like CSLLM, create "non-synthesizable" sets by selecting structures with low crystal-likeness scores from pre-trained models, but these may include actually synthesizable materials [16].

Temporal Validation Challenges

The ultimate test of synthesizability prediction is prospective validation—predicting which hypothetical materials will be successfully synthesized in the future. The SyntheFormer model addressed this through temporal splitting, training on data through 2018 and evaluating on materials reported from 2019-2025 [19]. This approach revealed that many thermodynamically stable candidates remain unsynthesized while some metastable compounds are successfully realized, demonstrating that stability alone is insufficient to predict experimental attainability [19].

Dataset Quality and Bias Concerns

Human-curated analysis reveals significant quality issues with automated text-mined datasets, with one study finding only 15% of identified outliers were correctly extracted in a text-mined dataset [11]. Additional biases include:

• Structural Bias: Models trained primarily on successfully synthesized materials may learn features correlated with historical synthesis preferences rather than fundamental synthesizability.
• Compositional Bias: Known materials databases overrepresent certain element combinations and underrepresent others, limiting model generalizability.
• Stability Proxy Limitation: Thermodynamic stability (energy above hull) captures only approximately 50% of synthesized materials, confirming its inadequacy as a sole synthesizability metric [6].

Evaluation of crystalline material synthesizability prediction models remains constrained by the fundamental lack of negative examples, making True Positive Rate the primary reliable metric while precision estimation requires indirect methods. Current state-of-the-art models achieve TPR values exceeding 92-95% for specific material domains through advanced architectures like teacher-student networks and domain-specific transfer learning. Emerging approaches using large language models and hierarchical transformers show promising results, with some claims of >98% accuracy on balanced datasets.

Future progress requires addressing several critical challenges: developing standardized temporal validation protocols, improving dataset quality through human curation, creating more sophisticated α-estimation techniques for precision approximation, and establishing domain-specific benchmarks. Most importantly, the field would benefit from increased reporting of failed synthesis attempts and development of shared resources documenting confirmed unsynthesizable compounds to alleviate the core PU learning constraint. Until then, researchers should interpret reported performance metrics with understanding of their inherent limitations and prioritize models demonstrating robust performance across multiple material classes and temporal validation schemes.

A Landscape of Modern Synthesizability Prediction Models and Their Metrics

The accurate prediction of crystalline material synthesizability represents a critical bottleneck in accelerating materials discovery. Conventional assessments relying on thermodynamic stability metrics, such as energy above the convex hull, often fail to capture the complex kinetic and experimental factors governing actual synthesis. This comparison guide evaluates two prominent computational approaches for synthesizability prediction: the established Crystal-Likeness Score (CLscore) utilizing graph convolutional networks with partially supervised learning, and emerging Graph Neural Network (GNN) architectures that directly learn from crystalline structures. We examine their performance characteristics, architectural implementations, and suitability for high-throughput virtual screening of novel materials.

Performance Comparison

The table below summarizes the key performance metrics and characteristics of CLscore and modern GNN-based approaches for crystal synthesizability prediction.

Table 1: Performance Comparison of Structure-Based Synthesizability Prediction Models

Metric	CLscore (PU Learning)	Modern GNN Variants	LLM-Based Approaches
Prediction Accuracy	87.4% (True Positive Rate) [20]	Consistently outperforms conventional GNNs [21]	98.6% (Synthesizability LLM) [5]
Primary Methodology	Positive-Unlabeled Learning with GCN [20]	Kolmogorov-Arnold Networks (KA-GNN), Fourier-based KAN layers [21]	Fine-tuned Large Language Models (CSLLM framework) [5]
Key Advantage	Captures structural motifs beyond thermodynamic stability [20]	Superior expressivity, parameter efficiency, and interpretability [21]	Direct prediction of synthesis methods and precursors [5]
Validation Performance	86.2% true positive rate for materials reported after training period [20]	Enhanced performance across 7 molecular benchmarks [21]	97.9% accuracy on complex structures with large unit cells [5]
Interpretability	Limited	Highlights chemically meaningful substructures [21]	Natural language explanations of synthesis pathways [5]

Table 2: Architectural Comparison of GNN Frameworks for Material Property Prediction

Architecture	Key Innovation	Application Domain	Performance
KA-GNN [21]	Integrates Kolmogorov-Arnold networks with Fourier-series-based functions	Molecular property prediction	Outperforms conventional GNNs in accuracy and efficiency [21]
MatGNet [22]	Mat2vec atomic embeddings with angular features via line graphs	Crystal property prediction	Surpasses previous models on JARVIS-DFT dataset [22]
ACES-GNN [23]	Explanation supervision for activity cliffs	Molecular activity prediction	Improves explainability and predictivity for activity cliffs [23]
GNN for Polycrystalline [24]	Microstructure graph embedding considering grain interactions	Polycrystalline material properties	~10% prediction error for magnetostriction across diverse microstructures [24]

Experimental Protocols & Methodologies

CLscore Implementation with PU Learning

The CLscore methodology employs a partially supervised learning approach to address the fundamental challenge in synthesizability prediction: the absence of verified negative examples in materials databases.

Dataset Construction: The model is trained on experimentally reported crystal structures from databases like the Materials Project as positive examples. The key innovation lies in treating unreported structures as unlabeled rather than negative, acknowledging they may include synthesizable materials not yet discovered [20].

Graph Convolutional Network Architecture: Crystal structures are represented as graphs where atoms form nodes and bonds form edges. The GCN classifier performs node embedding and graph-level representation learning through layer-wise propagation rules [20].

Training Procedure: The PU learning implementation uses a custom objective function that distinguishes confirmed synthesizable structures (positive) from unlabeled structures without assuming they are unsynthesizable. This avoids the false negative problem inherent in binary classification approaches [20].

CLscore Calculation: The model outputs a crystal-likeness score between 0 and 1, with scores >0.5 indicating high synthesizability probability. Validation showed 71 of the top 100 high-scoring virtual materials had indeed been previously synthesized [20].

KA-GNN Framework for Molecular Prediction

The Kolmogorov-Arnold Graph Neural Network represents a recent architectural innovation that integrates KAN modules into fundamental GNN components.

Fourier-Based KAN Layers: KA-GNN replaces traditional multilayer perceptrons with Fourier-series-based learnable activation functions. This enhancement allows the model to capture both low-frequency and high-frequency structural patterns in molecular graphs, providing stronger approximation capabilities for complex molecular functions [21].

Component Integration: KA-GNN systematically integrates KAN modules across three core GNN components: (1) node embedding initialization, (2) message passing operations, and (3) graph-level readout functions. This comprehensive replacement of conventional transformations creates a fully differentiable architecture with enhanced representational power [21].

Architectural Variants: The framework implements two specialized variants: KA-GCN (KAN-augmented Graph Convolutional Network) and KA-GAT (KAN-augmented Graph Attention Network). KA-GCN initializes node embeddings by processing atomic features and neighboring bond information through KAN layers, while KA-GAT additionally incorporates edge embeddings for more expressive representation learning [21].

Experimental Validation: On seven molecular benchmarks, KA-GNN consistently outperformed conventional GNNs in prediction accuracy and computational efficiency while providing improved interpretability through highlighting of chemically meaningful substructures [21].

Figure 1: KA-GNN Architecture Integrating Kolmogorov-Arnold Networks

Research Reagent Solutions

Table 3: Essential Computational Tools for Synthesizability Prediction Research

Tool Category	Specific Implementation	Research Application
Graph Neural Network Frameworks	GraphSAGE [25], MPNN [26], GCN [24], GAT [21]	Base architectures for structure-based learning
Materials Databases	Materials Project [20], ICSD [5], JARVIS-DFT [22]	Sources of experimental and computational crystal structures
Specialized Architectures	KA-GNN [21], MatGNet [22], ACES-GNN [23]	Domain-optimized models for specific prediction tasks
Interpretability Tools	Integrated Gradients [26], GNNExplainer [23]	Attribution methods for model explanations
Benchmarking Datasets	OGB [27], Molecular Benchmarks [21], Jarvis-DFT [22]	Standardized evaluation frameworks

Advanced GNN Applications in Materials Science

Specialized GNN Architectures

Beyond synthesizability prediction, GNN architectures have evolved to address specific challenges across materials science domains:

Polycrystalline Materials Modeling: A specialized GNN approach represents polycrystalline microstructures as graphs where each grain constitutes a node with features including Euler angles, grain size, and neighbor count. The adjacency matrix encodes physical contact relationships between grains. This model achieved approximately 10% prediction error for magnetostriction in Tb₀.₃Dy₀.₇Fe₂ alloys while quantifying feature importance at the individual grain level [24].

Reaction Yield Prediction: Comparative studies of GNN architectures for chemical reaction yield prediction identified Message Passing Neural Networks (MPNN) as the top performer (R²=0.75) across diverse cross-coupling reactions. Integrated gradients methods provided interpretable insights into descriptor contributions, highlighting the potential for explainable reaction optimization [26].

Activity Cliff Explanation: The ACES-GNN framework addresses the "black-box" limitation of conventional models by incorporating explanation supervision for activity cliffs—structurally similar molecules with significant potency differences. This approach improves both prediction accuracy and attribution quality by aligning model reasoning with chemist intuition [23].

Figure 2: Comprehensive Workflow for Crystal Synthesizability Assessment

Performance Optimization Techniques

Recent advances in GNN methodologies have addressed specific performance limitations:

Label Propagation Enhancement: The Label as Equilibrium approach resolves over-fitting issues in label reuse for node classification by implementing supervision concealment and infinite iterations with constant memory consumption. This technique boosted prevailing GNN accuracy by 2.31% on average, demonstrating significant potential for materials classification tasks [27].

Angular Feature Incorporation: MatGNet's integration of angular features through line graphs and mat2vec embeddings significantly improved crystal property prediction accuracy beyond traditional GCN approaches, though with increased computational overhead. This tradeoff between expressive power and efficiency remains a key consideration for large-scale virtual screening [22].

The evolving landscape of structure-based synthesizability prediction demonstrates a clear trajectory from descriptor-based machine learning to specialized deep learning architectures. While CLscore established the viability of GCN-based approaches with PU learning for synthesizability screening, modern GNN variants like KA-GNN offer enhanced accuracy, efficiency, and interpretability. The emerging paradigm integrates these architectures into comprehensive frameworks that simultaneously predict synthesizability, identify synthetic routes, and suggest appropriate precursors. For research applications, the selection between these approaches involves tradeoffs between interpretability (CLscore), predictive accuracy (KA-GNN), and comprehensive synthesis planning (CSLLM). As these methodologies mature, they promise to significantly reduce the experimental burden in materials discovery by providing reliable synthesizability assessments before resource-intensive synthesis attempts.

The evaluation of crystalline material synthesizability has long been a critical bottleneck in materials science and drug development. Traditional prediction methods relying on thermodynamic formation energy (Ehull) or phonon spectrum analysis have faced significant limitations, often failing to bridge the gap between theoretical design and experimental synthesis. The emergence of Large Language Models (LLMs) represents a paradigm shift in this field, moving beyond textual understanding to achieve unprecedented accuracy in predicting which theoretical materials can be successfully synthesized. This transformation is particularly evident in pharmaceutical development, where crystal structure prediction directly impacts drug stability, bioavailability, and intellectual property protection.

Specialized LLMs are now demonstrating remarkable capabilities in accurately predicting material synthesizability and properties. The CSLLM (Crystal Synthesis Large Language Models) framework exemplifies this progress, achieving a 98.6% prediction accuracy for crystalline material synthesizability, substantially outperforming traditional computational methods that often struggle with practical synthesis feasibility [28]. This breakthrough performance stems from innovative approaches to representing and processing materials data as textual representations that LLMs can effectively analyze.

Comparative Analysis: LLMs Versus Traditional Methods

Quantitative Performance Comparison

The table below summarizes the performance differences between LLM-based approaches and traditional computational methods for predicting crystalline material synthesizability:

Prediction Method	Accuracy Rate	Key Strengths	Primary Limitations
Synthesizability LLM (CSLLM)	98.6% [28]	Exceptional accuracy for complex structures; strong generalization to large-unit-cell structures (97.8% accuracy) [28]	Requires balanced training datasets; dependent on quality text representations
Thermodynamic (Ehull ≥0.1eV/atom)	74.1% [28]	Established physical principles; interpretable results	Frequently misclassifies synthesizable metastable materials
Phonon Frequency (≥-0.1THz)	82.2% [28]	Identifies dynamical instabilities	Limited practical predictive value for synthesizability
Crystal Structure Prediction (CSP)	Varies by complexity [29]	Physics-based; comprehensive conformational sampling	Computationally intensive; accuracy decreases with molecular complexity

Case Study: Complex Pharmaceutical Compounds

In rigorous blind tests conducted by the Cambridge Crystallographic Data Centre (CCDC), traditional CSP methods struggled with complex pharmaceutical compounds like Pfizer's Alzheimer candidate drug PD-0118057 (43 atoms, 7 flexible dihedral angles). While the best traditional approaches identified 4 of 5 known crystal forms, LLM-enhanced methods successfully predicted all 5 experimental polymorphs with high structural accuracy (RMS distances of 0.2Å to 0.4Å) [29]. For the challenging ROY (5-methyl-2-[(2-nitrophenyl)-amino]-3-thiophenecarbonitrile) system with 12 known polymorphs, LLM-augmented approaches correctly identified all experimental structures, outperforming traditional methods that found only 7-10 forms [29].

Experimental Protocols and Methodologies

CSLLM Framework Architecture

The exceptional performance of LLMs in crystalline materials prediction stems from specialized frameworks and methodologies. The CSLLM framework employs three dedicated models fine-tuned from LLaMA3-8B using Low-Rank Adaptation (LoRA): Synthesizability LLM for synthesizability prediction, Method LLM for synthesis route classification (97.98% accuracy), and Precursor LLM for precursor identification (>90% success rate) [28].

The critical innovation enabling LLM application to materials science is the "material string" text representation, which compresses conventional Crystallographic Information File (CIF) data by 94% into a 102-character string format [28]. This representation systematically encodes:

• Space group numbers (e.g., 221)
• Lattice constants (e.g., 3.897Å, 3.897Å, 3.897Å, 90.0°, 90.0°, 90.0°)
• Atomic symbols and Wyckoff positions (e.g., (Ca-1a[0.0,0.0,0.0])→(Ti-1b[0.5,0.5,0.5])→(O-3c[0.5,0.5,0.5]))

This textual representation allows the LLM to process crystal structures with the same techniques used for natural language, while maintaining all essential structural information needed for accurate prediction.

Dataset Construction and Training

The CSLLM framework was trained on a meticulously balanced dataset containing 150,120 materials, including 70,120 experimentally confirmed structures from the Inorganic Crystal Structure Database (ICSD) as positive samples and 80,000 theoretical structures carefully selected through Positive-Unlabeled (PU) learning as negative samples [28]. This comprehensive dataset covers seven crystal systems and compounds containing 1-7 elements with atomic numbers ranging from 1-94, ensuring broad coverage of chemical space.

To address the critical challenge of LLM "hallucination" in scientific applications, researchers implemented rigorous validation protocols. Ten repeated tests demonstrated minimal prediction variance (<0.06% difference rate), ensuring highly reproducible results essential for scientific and pharmaceutical applications [28].

Advanced LLM Architectures for Scientific Accuracy

DRAG: Enhancing Retrieval Accuracy for Scientific Domains

While early LLM applications in science faced challenges with factual accuracy, specialized architectures like DRAG (Lexical Diversity-aware Retrieval Augmented Generation) have demonstrated substantial improvements. DRAG addresses the vocabulary diversity problem in scientific domains where the same concept may be described using different terminology (e.g., "profession" vs. "occupation" vs. "career") [30].

The DRAG framework employs two innovative components. The Diversity-sensitive Relevance Analyzer (DRA) classifies query terms into "invariant," "variant," and "supplementary" components with different matching strategies [30]. The Risk-guided Sparse Calibration (RSC) strategy then identifies and calibrates only high-risk tokens during generation, minimizing computational overhead while maximizing accuracy [30].

In rigorous testing, DRAG increased factual accuracy by 45.5% compared to base LLMs and outperformed the next best RAG method by 4.9% on the PopQA dataset [30]. For complex multi-hop reasoning tasks (HotpotQA), DRAG's advantage increased to 10.6% over alternative methods, demonstrating particularly strong performance for complex scientific queries [30].

Specialized Scientific LLMs Beyond General-Purpose Models

The scientific community has developed specialized LLMs tailored to specific research domains, moving beyond general-purpose models like GPT and Llama. In materials science, models like LLaMat excel at material-specific natural language processing and crystal structure generation [31]. SurFF (Surface Foundation Model) represents another specialized approach, using equivariant graph neural networks to predict surface energy and morphology across intermetallic crystals with DFT-level accuracy (3 meV/Ų error) but with 10⁵-fold acceleration [32].

These specialized models typically employ techniques like retrieval-augmented generation (RAG) and chain-of-thought (CoT) prompting to enhance scientific reasoning [31]. Multi-agent systems with LLMs assuming different roles (researcher, reviewer, moderator) further mimic collaborative human scientific reasoning for hypothesis generation and validation [31].

Key Research Reagent Solutions for LLM-Enhanced Materials Prediction

The table below outlines essential computational tools and data resources for implementing LLM-based crystalline materials prediction:

Tool/Resource	Function	Application Context
CSLLM Framework	Predicts crystal synthesizability and precursors	Open-source interactive interface for CIF/POSCAR file input [28]
Cambridge Structural Database (CSD)	Reference database of experimental crystal structures	Training data source; validation benchmark for predictions [29]
Materials Project Database	Repository of computed materials properties	Source of theoretical structures for training and validation [28]
LoRA (Low-Rank Adaptation)	Efficient LLM fine-tuning method	Adapting foundation LLMs to specialized materials science tasks [28]
vLLM with PagedAttention	High-throughput LLM inference framework	Deployment of materials prediction models with optimized memory usage [33]
DRAG Architecture	Enhanced retrieval for scientific vocabulary	Improving factual accuracy in materials science literature analysis [30]

The integration of LLMs into crystalline materials research represents more than an incremental improvement—it constitutes a fundamental transformation in how scientists approach synthesizability prediction. By achieving 98.6% prediction accuracy and successfully identifying 45,632 synthesizable materials from theoretical candidates, LLMs have dramatically accelerated the materials discovery pipeline [28]. The emerging generation of scientific LLMs operates not merely as pattern recognition systems but as sophisticated reasoning engines that combine textual understanding with domain-specific knowledge.

As these technologies continue evolving, they promise to further close the gap between theoretical materials design and experimental synthesis. The development of interactive platforms that accept standard crystallographic file formats makes this technology increasingly accessible to materials scientists and pharmaceutical researchers worldwide [28]. With LLMs now capable of not only predicting synthesizability but also recommending specific synthesis methods and precursors with >90% success rates [28], we are witnessing the emergence of a new paradigm in materials research—one where AI-powered prediction and human expertise collaboratively advance the frontiers of materials science and drug development.

In the field of computational materials discovery, accurately predicting which theoretical crystalline materials can be successfully synthesized in a laboratory is a fundamental challenge. The concept of synthesizability extends beyond mere thermodynamic stability to encompass whether a material is synthetically accessible with current experimental capabilities, a critical filter for prioritizing candidates from vast computational databases [6]. Among the various computational approaches developed, composition-only models represent a distinct category that relies solely on a material's chemical formula to predict synthesizability. This guide provides an objective comparison of these models, examining their performance against more complex alternatives and analyzing the trade-offs between their predictive accuracy and practical utility within research workflows.

Performance Comparison of Synthesizability Prediction Methods

Composition-only models occupy a specific niche in the synthesizability prediction landscape. They can be deployed early in the discovery pipeline when only chemical formulas are known, offering computational efficiency but with inherent limitations in predictive power compared to structure-aware approaches. The table below summarizes the key characteristics and performance metrics of major synthesizability prediction methods, including composition-only models and their more advanced counterparts.

Table 1: Comparative Analysis of Synthesizability Prediction Methods

Method Name	Model Type	Input Data	Key Performance Metrics	Primary Advantages	Primary Limitations
SynthNN [6]	Composition-only (Deep Learning)	Chemical composition	7× higher precision than DFT formation energies; outperformed human experts by 1.5× precision	Computationally efficient; requires only chemical formula; rapid screening of vast composition spaces	Cannot differentiate between polymorphs; limited accuracy for complex compositions
Charge-Balancing [6]	Heuristic/Rule-based	Chemical composition	Only 37% of known synthesized materials are charge-balanced; 23% for binary cesium compounds	Chemically intuitive; computationally inexpensive	Inflexible constraint; poor performance as standalone synthesizability predictor
CSLLM [5]	Structure-aware (Large Language Model)	Crystal structure (text representation)	98.6% accuracy in synthesizability prediction; significantly outperforms thermodynamic (74.1%) and kinetic (82.2%) stability methods	High accuracy; can also predict synthetic methods and precursors (>90% accuracy)	Requires structural information; computationally intensive
Unified Composition-Structure Model [3]	Hybrid (Composition + Structure)	Both composition and crystal structure	Successfully guided experimental synthesis of 7 out of 16 target compounds, including novel materials	Integrates complementary signals from composition and structure; demonstrated experimental validation	Requires complete structural data; more complex implementation

Experimental Protocols and Methodologies

Composition-Only Model Development (SynthNN)

The development of SynthNN exemplifies the composition-only approach to synthesizability prediction [6]. The experimental protocol involves several methodical stages:

Data Curation and Representation:

• Training Data Source: Models are trained on the Inorganic Crystal Structure Database (ICSD), which contains historically synthesized inorganic crystalline materials [6].
• Input Representation: The atom2vec framework represents chemical formulas through a learned atom embedding matrix optimized alongside other neural network parameters [6]. This approach learns optimal representations directly from the distribution of synthesized materials without requiring pre-defined chemical descriptors.
• Handling Unlabeled Data: A critical challenge is the lack of confirmed "unsynthesizable" examples in scientific literature. Researchers address this through Positive-Unlabeled (PU) learning algorithms, treating artificially generated materials as unlabeled data and probabilistically reweighting them according to their likelihood of being synthesizable [6].

Model Architecture and Training:

• The model employs a deep learning architecture where the dimensionality of the composition representation is treated as a hyperparameter optimized before training.
• Training utilizes a semi-supervised approach that accounts for the incomplete labeling of artificially generated examples, with the ratio of artificial to synthesized formulas being a key hyperparameter.
• The model learns chemical principles such as charge-balancing, chemical family relationships, and ionicity directly from data without explicit programming of these rules [6].

Benchmarking Protocol:

• Performance is evaluated against baseline methods including random guessing and charge-balancing approaches.
• Standard classification metrics are calculated by treating synthesized materials as positive examples and artificially generated materials as negative examples.
• Due to the PU learning framework, F1-score is often emphasized as a key evaluation metric alongside precision and recall [6].

Comparative Evaluation Framework

Rigorous evaluation of composition-only models requires comparison against multiple alternative approaches:

Performance Against Human Experts:

• In head-to-head material discovery comparisons, SynthNN outperformed all 20 expert material scientists, achieving 1.5× higher precision and completing the task five orders of magnitude faster than the best human expert [6].

Comparison with Traditional Computational Methods:

• Composition-only models demonstrate 7× higher precision in identifying synthesizable materials compared to traditional DFT-calculated formation energies [6].
• They significantly outperform charge-balancing approaches, which fail to predict synthesizability for many known compounds due to the inflexibility of the charge neutrality constraint [6].

Limitations Assessment:

• A fundamental limitation is evaluated: composition-only models cannot differentiate between different crystal structures (polymorphs) of the same chemical composition [6].
• This limitation becomes critical for materials like carbon (diamond vs. graphite) where the same composition yields materials with drastically different properties and synthesizability [34].

The Scientist's Toolkit: Essential Research Reagents

Table 2: Key Computational Tools and Data Resources for Synthesizability Prediction Research

Tool/Resource	Type	Primary Function in Research	Access Considerations
Inorganic Crystal Structure Database (ICSD) [6] [5]	Database	Provides curated data on experimentally synthesized inorganic crystalline structures for model training and validation	Subscription-based access; comprehensive but requires licensing
Materials Project [3]	Database	Source of computationally predicted structures with DFT-calculated properties; used for benchmarking and negative sample generation	Freely accessible; API available for automated data retrieval
atom2vec [6]	Algorithm	Learns optimal compositional representations directly from data without requiring pre-defined chemical descriptors	Implementation-dependent; requires programming expertise
Positive-Unlabeled Learning [6]	Machine Learning Framework	Handles the lack of confirmed negative examples by treating un synthesized materials as unlabeled data	Specialized implementation needed beyond standard classification
Wyckoff Encode [34]	Structural Descriptor	Captures symmetry information in crystal structures for structure-based models; not used in composition-only approaches	Openly available in some research codebases

Workflow Diagram: Composition-Only Synthesizability Prediction

The following diagram illustrates the standard workflow for developing and applying composition-only synthesizability prediction models, highlighting both their streamlined nature and inherent limitations compared to more comprehensive approaches.

Diagram 1: Composition-Only Model Workflow and Limitation

Composition-only models represent a pragmatic trade-off in the synthesizability prediction landscape. Their principal advantage lies in computational efficiency and applicability during early discovery stages when only compositional information is available. The experimental success of models like SynthNN demonstrates they can significantly outperform traditional DFT-based approaches and even human experts in specific screening tasks [6]. However, their fundamental limitation in differentiating polymorphs constrains their utility for final candidate selection [34].

The choice between composition-only and more complex structure-aware models depends on the research context. For initial high-throughput screening of vast compositional spaces, composition-only models provide an efficient filtering mechanism. For final candidate prioritization and synthesis planning, structure-aware approaches like CSLLM [5] or hybrid models [3] offer superior accuracy despite greater computational demands. As materials informatics evolves, the strategic integration of both approaches—using composition-only models for initial screening followed by structure-aware validation—represents the most promising path toward accelerating experimental materials discovery.

The accurate prediction of crystalline material synthesizability represents a central challenge in accelerating the discovery of new materials for pharmaceuticals, electronics, and energy applications. Traditional approaches often rely on单一的 (single) descriptors, such as those derived solely from composition or structure, leading to incomplete predictive models. This comparison guide evaluates three advanced computational frameworks that integrate both compositional and structural signals to overcome these limitations. By systematically examining their architectures, experimental protocols, and performance metrics, this analysis aims to inform researchers and development professionals about the current state-of-the-art and its practical implications for materials design within the broader context of accuracy metrics for synthesizability prediction research.

Framework Comparison at a Glance

The table below provides a high-level comparison of three prominent integrated frameworks for crystalline material property prediction, highlighting their core approaches and performance.

Table 1: Overview of Integrated Frameworks for Crystal Property Prediction

Framework Name	Core Integration Method	Primary Prediction Tasks	Reported Performance Highlights
LLM-Prop [35]	Fine-tuned encoder of a transformer model (T5) on text descriptions of crystals.	Band gap, formation energy, unit cell volume.	≈8% improvement on band gap prediction over GNN baselines [35].
CSLLM [16]	Three specialized LLMs fine-tuned on a comprehensive "material string" representation.	Synthesizability, synthetic methods, suitable precursors.	98.6% accuracy in synthesizability prediction [16].
Language Representation Framework [36]	Pretrained transformer models (MatBERT, MatSciBERT) for contextual embeddings of material text.	Material similarity, multi-property ranking (e.g., for thermoelectrics).	Effective recall of relevant candidates and property prediction comparable to specialized models [36].

Detailed Performance and Experimental Data

A deeper examination of the quantitative results and experimental setups reveals the distinct advantages of each framework under specific conditions.

Table 2: Detailed Quantitative Performance Metrics

Framework	Key Experimental Results	Comparative Baseline Performance	Dataset Used
LLM-Prop [35]	- 65% improvement on unit cell volume prediction vs. GNNs.- Comparable performance on formation energy/atom.	Outperforms ALIGNN (state-of-the-art GNN) and fine-tuned MatBERT (with fewer parameters) [35].	TextEdge (curated benchmark with crystal text descriptions) [35].
CSLLM [16]	- 98.6% synthesizability prediction accuracy on test data.>90% accuracy for synthetic method classification.>80% success for precursor prediction.	Significantly outperforms thermodynamic (74.1%) and kinetic (82.2%) stability methods [16].	Balanced dataset of 70,120 synthesizable (ICSD) and 80,000 non-synthesizable structures [16].
Language Representation Framework [36]	- 94 out of 100 high-zT materials showed statistically significant recall.- Effective identification of under-explored material spaces with high predicted performance.	Language-based similarity recall shows distinct advantage over baseline representations (Mat2Vec, fingerprints) and random sampling [36].	116,000 materials from various sources; text descriptions generated by Robocrystallographer [36].

Experimental Protocols and Methodologies

LLM-Prop Protocol

The LLM-Prop framework leverages the encoder of a T5 transformer model. The key methodological steps involve specific input preprocessing to adapt crystal text descriptions for the language model [35]:

• Input Preprocessing: Stop words are removed, while digits and signs potentially carrying critical information are retained.
• Numerical Tokenization: Bond distances and angles, along with their units, are replaced with special tokens [NUM] and [ANG] to compress the sequence length and mitigate LLMs' known weaknesses in numerical reasoning.
• Classification Token: A [CLS] token is prepended to the input sequence. The final hidden state corresponding to this token is used as the aggregate representation for the regression or classification task, following the practice established in BERT models [35].
• Fine-tuning: The T5 encoder is fine-tuned on the preprocessed text descriptions, with a linear layer added on top for the final prediction task.

CSLLM Protocol

The Crystal Synthesis LLM (CSLLM) framework employs a multi-model approach, with each LLM specialized for a distinct subtask. The core experimental methodology is [16]:

• Dataset Curation:
  • Positive Samples: 70,120 synthesizable crystal structures were meticulously selected from the Inorganic Crystal Structure Database (ICSD).
  • Negative Samples: 80,000 non-synthesizable structures were identified by applying a pre-trained Positive-Unlabeled (PU) learning model to a pool of over 1.4 million theoretical structures and selecting those with the lowest CLscore (a synthesizability score) [16].
• Text Representation: A "material string" format was developed to convert crystal structures into a concise, reversible text description that efficiently encapsulates lattice parameters, composition, atomic coordinates, and symmetry information without redundancy [16].
• Model Fine-tuning: Three separate LLMs were fine-tuned on this balanced dataset using the material string representation for the specific tasks of synthesizability prediction, synthetic method classification, and precursor identification.

Language Representation Framework for Exploration

This framework focuses on materials exploration and recommendation using a funnel-based architecture, which consists of a recall step followed by a ranking step [36]:

• Representation Generation:
  • Compositional: Material formulae (e.g., "PbTe") are embedded using models like Mat2Vec or contextual embeddings from MatBERT/MatSciBERT.
  • Structural: Automated text descriptions of crystal structures (e.g., "PbTe is Halite, Rock Salt structured...") generated by Robocrystallographer are embedded using the same BERT models [36].
• Recall Step: For a given query material, candidate materials are generated by calculating cosine similarity between the query's language representation and all materials in the database in the shared embedding space.
• Ranking Step: Recalled candidates are evaluated and ranked using a multi-task learning model (a Multi-gate Mixture-of-Experts, or MMoE) that predicts multiple target properties simultaneously, leveraging correlations between tasks to improve accuracy [36].

Workflow and Signaling Pathways

The following diagram illustrates the logical workflow of a generic integrated framework that combines compositional and structural signals for property prediction, synthesizing common elements from the discussed methodologies.

Integrated Framework Workflow

For researchers aiming to implement or benchmark these integrated frameworks, the following computational "reagents" and resources are critical.

Table 3: Key Resources for Integrated Framework Research

Resource Name/Type	Function in Research	Relevance to Integrated Frameworks
TextEdge Dataset [35]	A benchmark dataset pairing crystal text descriptions with their properties.	Serves as a public benchmark for training and evaluating text-based models like LLM-Prop.
Balanced Synthesizability Dataset [16]	A curated set of ~150k known synthesizable and non-synthesizable structures.	Essential for training high-fidelity synthesizability predictors like CSLLM, mitigating data bias.
Robocrystallographer [36]	A tool that generates human-readable text descriptions from crystal structures.	Automatically creates the structural text input required by language representation models.
MatBERT / MatSciBERT [36]	Domain-specific language models pre-trained on materials science literature.	Provide foundational, context-aware embeddings that capture domain knowledge for composition and structure.
Universal Model for Atoms (UMA) [37]	A machine learning interatomic potential trained across diverse chemical domains.	Enables fast and accurate relaxation and ranking of crystal structures, as used in the FastCSP workflow.

Specialized Models for Solid-State and Perovskite Synthesis

The discovery and synthesis of novel crystalline materials, particularly perovskites for energy and optoelectronic applications, represent a critical frontier in materials science [38]. However, a significant bottleneck persists: the transition from theoretical prediction to experimental realization. For years, researchers have relied on computational proxies like thermodynamic stability (energy above the convex hull) or kinetic stability (phonon spectra) to screen for synthesizable materials [5] [11]. Unfortunately, these metrics are imperfect; numerous metastable structures are synthesizable, while many thermodynamically favorable ones are not [5]. This gap has spurred the development of specialized data-driven models that learn the complex patterns of synthesizability directly from experimental data, offering a more direct and accurate guide for experimentalists [6]. This guide objectively compares the performance, methodologies, and applications of the latest generation of synthesizability prediction models, framing them within the critical context of accuracy metrics for crystalline material research.

Comparative Analysis of Model Performance

The performance of synthesizability prediction models is typically evaluated using metrics such as accuracy, precision, and F1-score on held-out test sets. The table below provides a quantitative comparison of contemporary models.

Table 1: Performance Metrics of Specialized Synthesizability Prediction Models

Model Name	Primary Scope	Reported Accuracy	Key Performance Highlights	Key Advantages
Crystal Synthesis LLM (CSLLM) [5]	General 3D Crystal Structures	98.6%	Outperforms thermodynamic (74.1%) and kinetic (82.2%) stability screening; 97.9% accuracy on complex structures [5].	Predicts synthesizability, synthetic methods, and precursors; exceptional generalization.
SynthNN [6]	Inorganic Crystalline Compositions	Not Specified	7x higher precision in identifying synthesizable materials than DFT-calculated formation energies [6].	Requires only chemical formulas, no structural data needed; high computational efficiency.
Positive-Unlabeled (PU) Learning Model (Jang et al.) [5] [11]	General 3D Crystals / Ternary Oxides	87.9% [5] to 92.9% [11]	Used to generate negative samples for training other models like CSLLM [5].	Effective for semi-supervised learning with limited negative data.
Question Answering (QA) MatSciBERT [39]	Information Extraction (e.g., Bandgaps)	N/A (Extraction Task)	Achieved a 61.3 F1-score for extracting material-property relationships from text, outperforming other NLP tools [39].	Extracts precise data from scientific literature; reduces "hallucination" common in generative models.

Detailed Model Methodologies and Experimental Protocols

Understanding the experimental and computational protocols behind these models is crucial for assessing their reliability and applicability.

The Crystal Synthesis Large Language Model (CSLLM) Framework

The CSLLM represents a groundbreaking approach that uses three specialized LLMs to address the synthesis prediction pipeline [5].

• Dataset Construction: The model was trained on a balanced dataset of 70,120 synthesizable crystal structures from the Inorganic Crystal Structure Database (ICSD) and 80,000 non-synthesizable structures identified from over 1.4 million theoretical structures using a pre-trained PU learning model [5].
• Text Representation (Material String): A key innovation was the development of a concise text representation for crystal structures, termed "material string." This format efficiently encodes space group, lattice parameters, atomic species, Wyckoff positions, and fractional coordinates, making it suitable for LLM processing [5].
• Model Fine-Tuning: The framework involves three fine-tuned LLMs:
  • Synthesizability LLM: Predicts whether an arbitrary 3D crystal structure is synthesizable.
  • Method LLM: Classifies the likely synthetic method (e.g., solid-state or solution).
  • Precursor LLM: Identifies suitable solid-state synthesis precursors for binary and ternary compounds [5].
• Validation: Model performance was validated through standard train-test splits and demonstrated exceptional generalization on structures with complexity far exceeding the training data [5].

Data-Driven Workflow for Solid-State Synthesis Planning

Complementing the CSLLM, a more chemistry-focused workflow has been developed for planning solid-state synthesis reactions, emphasizing thermodynamic selectivity [40].

• Primary and Secondary Competition Metrics: This approach introduces two novel metrics derived from thermodynamic data. The Primary Competition metric gauges the favorability of the target product forming versus competing compounds from the original precursors. The Secondary Competition metric assesses the stability of the target product against decomposition into unwanted side products after its formation [40].
• Data Source: The workflow utilizes a large thermodynamic database, such as the Materials Project, to calculate the reaction energies for thousands of potential synthesis pathways [40].
• Experimental Protocol: In a case study on barium titanate (BaTiO₃) synthesis, the model identified 82,985 possible reactions. Nine were selected for experimental testing, including reactions involving unconventional precursors like barium sulfide (BaS) and sodium metatitanate (Na₂TiO₃). The reactions were characterized using techniques like synchrotron powder X-ray diffraction to track phase formation and impurity levels [40].
• Validation: The experimentally observed formation of the target material and impurities showed strong correlation with the predicted primary and secondary competition metrics, validating the approach [40].

Table 2: Essential Research Reagents and Solutions for Solid-State Synthesis

Reagent / Material	Function in Synthesis	Application Example
Conventional Precursors (e.g., BaCO₃, TiO₂)	Source of cationic components for the target material.	Conventional synthesis of BaTiO₃ [40].
Unconventional Precursors (e.g., BaS, BaCl₂, Na₂TiO₃)	Can offer kinetic or thermodynamic pathways that lower impurity formation.	Alternative, more efficient synthesis routes for BaTiO₃ [40].
Solid-State Synthesis Dataset (e.g., TMR dataset)	Provides text-mined data on heating temperatures, times, and precursors from literature to train machine learning models [41].	Used to train models that predict optimal synthesis conditions [41].

The following diagram illustrates the logical workflow of the CSLLM framework, from data preparation to final prediction.

Graph 1: CSLLM Framework Workflow. This diagram outlines the three-stage process of the Crystal Synthesis Large Language Model framework, from curating crystal structure data to generating synthesizability, method, and precursor predictions.

Discussion and Future Perspectives

The advent of specialized models like CSLLM and data-driven workflows marks a significant leap beyond traditional heuristic or stability-based screening. The key insight is that synthesizability is a complex property that can be learned from the collective record of successful syntheses. The high accuracy of these models, as shown in Table 1, demonstrates their potential to dramatically reduce failed synthetic attempts.

Future developments will likely focus on integrating kinetic factors more explicitly, as precursor properties (e.g., melting points) are already known to be strong predictors of optimal solid-state reaction temperatures [41]. Furthermore, expanding the scope of models to include more detailed reaction conditions, such as atmosphere and pressure, and applying them to a broader range of material classes, including the diverse family of lead-free perovskites [38], will be crucial. As these tools become more sophisticated and user-friendly, they are poised to become an indispensable part of the materials researcher's toolkit, accelerating the rational design and discovery of next-generation functional materials.

Optimizing Predictive Performance: Data, Design, and Explainability

The rise of data-driven science represents a fourth paradigm in materials research, following historical eras of experimental, theoretical, and computational discovery [42]. In this new paradigm, the quality and nature of training data fundamentally constrain the accuracy of predictive models, especially for complex challenges like forecasting crystalline material synthesizability. Two primary approaches have emerged for constructing these essential datasets: human-curated data, characterized by expert validation, and text-mined data, extracted automatically from scientific literature using Natural Language Processing (NLP) [43]. The selection between these data types involves critical trade-offs between precision, coverage, and scalability, directly impacting the performance of subsequent machine learning applications. This guide objectively compares these methodologies within the specific context of developing accurate synthesizability predictors, providing researchers with evidence-based insights for selecting appropriate data strategies for their discovery pipelines.

Fundamental Definitions and Data Characteristics

Human-Curated Data

Human-curated data consists of information that has been carefully selected and organized by experts in the field. This data type is typically well-established and has undergone thorough validation [43]. In materials science, prominent sources of curated data include specialized databases such as the Inorganic Crystal Structure Database (ICSD), which provides a comprehensive collection of experimentally synthesized crystalline structures used for training synthesizability models like SynthNN and CSLLM [6] [16]. The curation process imposes a high degree of veracity, making this data type particularly valuable for benchmarking and validating fundamental material properties.

Text-Mined Data

Text-mined data is information extracted automatically from scientific literature using high-performance Natural Language Processing (NLP) tools [43]. This approach can process millions of full-text articles to identify material associations, synthesis parameters, and property data that would be infeasible to collect manually [44]. While potentially less established than curated data, text mining offers a powerful source of novel insights and can capture the collective knowledge embedded in the vast, unstructured corpus of published research [43] [45]. Benchmark studies confirm that text mining of full-text articles consistently yields more associations and higher accuracy compared to using only abstracts [44].

Table 1: Core Characteristics of Human-Curated vs. Text-Mined Data

Characteristic	Human-Curated Data	Text-Mined Data
Fundamental Definition	Expert-validated information from trusted sources [43]	NLP-extracted information from scientific literature [43]
Primary Sources	CLINGEN, ClinVar, UniProt, ICSD [43] [6]	Full-text scientific articles from Elsevier, Springer, PMC [44]
Verification Process	Thorough human validation	Automated extraction with potential manual review
Inherent Advantages	High accuracy, established knowledge	Broad coverage, novel relationship discovery
Typical Applications	Model training benchmarks, stability prediction	Knowledge graph construction, precursor identification

Quantitative Performance Comparison in Synthesizability Prediction

The ultimate test for any materials data strategy lies in its performance when deployed within machine learning workflows. The following comparative analysis examines how models trained on these different data paradigms perform on the critical task of crystalline material synthesizability prediction.

Table 2: Performance Comparison of Synthesizability Prediction Models

Model (Data Source)	Data Foundation	Prediction Accuracy	Key Performance Metrics
CSLLM Framework [16]	Curated data from ICSD	98.6%	State-of-the-art accuracy on testing data
SynthNN [6]	Curated data from ICSD	7× higher precision than DFT formation energy	1.5× higher precision than best human expert
CPUL Model [46]	Positive-unlabeled learning from MP database	93.95%	True positive prediction accuracy
Traditional DFT Screening [16]	Computed formation energies	74.1% (Energy above hull ≥0.1 eV/atom)	Thermodynamic stability metric
Kinetic Stability Screening [16]	Phonon spectrum analysis	82.2% (Lowest frequency ≥ -0.1 THz)	Kinetic stability metric

The quantitative evidence demonstrates a clear performance hierarchy. Models like the Crystal Synthesis Large Language Models (CSLLM) framework, trained on expertly curated data from the ICSD, achieve remarkable accuracy up to 98.6% in distinguishing synthesizable from non-synthesizable crystal structures [16]. Similarly, the SynthNN model demonstrates 7× higher precision in identifying synthesizable materials compared to traditional screening using DFT-calculated formation energies [6]. These results significantly outperform conventional physics-based screening methods that rely solely on thermodynamic or kinetic stability metrics [16].

The superior performance of models trained on curated data stems from their foundation in experimentally verified material records. The CSLLM framework, for instance, was trained on a balanced dataset containing 70,120 synthesizable crystal structures from the ICSD alongside 80,000 non-synthesizable structures identified through positive-unlabeled learning [16]. This careful data construction enables the model to learn the complex chemical principles governing synthesizability—including charge-balancing, chemical family relationships, and ionicity—directly from the distribution of realized materials [6].

Experimental Protocols and Methodologies

Curated Data Workflow for Synthesizability Prediction

The application of human-curated data follows a structured experimental pathway designed to maximize data quality and model reliability.

The workflow for utilizing curated data begins with experimental synthesis and characterization of materials, followed by expert entry into structured databases like the ICSD [6] [16]. For model training, relevant features are extracted from these verified records. In the SynthNN approach, this involves using an atom2vec representation that learns optimal features directly from the distribution of synthesized materials without requiring prior chemical assumptions [6]. The CSLLM framework employs a specialized "material string" text representation that integrates essential crystal information in a format suitable for large language model processing [16]. Models are then trained and evaluated against held-out test sets, with performance validated through metrics like accuracy, precision, and comparison against human experts or traditional methods [6] [16].

Text Mining Methodology for Knowledge Extraction

Text mining operationalizes the vast knowledge embedded in scientific literature through a multi-stage processing pipeline.

The text mining pipeline processes massive collections of scientific documents—with studies analyzing up to 15 million full-text articles [44]. After collection, documents undergo preprocessing including PDF-to-text conversion, language detection (filtering for English content), and cleanup of non-printable characters or poorly converted text [44]. The core extraction phase employs Named Entity Recognition (NER) systems to identify relevant materials science concepts (materials, properties, methods) followed by relationship extraction to establish connections between these entities [44]. The extracted information is then integrated into structured databases or knowledge graphs that support downstream applications such as synthesis planning and precursor identification [44] [16]. Benchmark studies demonstrate that full-text mining consistently outperforms abstract-only approaches, extracting more complete information with higher accuracy [44].

Table 3: Key Experimental Resources for Synthesizability Research

Resource/Solution	Type	Primary Function	Exemplary Use Case
ICSD Database [6] [16]	Curated Data	Provides experimentally verified crystal structures	Positive examples for synthesizability model training
Materials Project API [46]	Computational Data	Access to DFT-calculated material properties	Source of hypothetical structures for negative examples
Atom2Vec Representation [6]	Computational Tool	Learns optimal material representations from data	Feature extraction in SynthNN without chemical assumptions
Material String Format [16]	Data Representation	Text-based crystal structure encoding	LLM-friendly input for CSLLM framework
Positive-Unlabeled Learning [6] [46]	ML Methodology	Handles lack of verified negative examples	Estimating synthesizability probability (CLscore)
NER Systems [44]	Text Mining Tool	Extracts material entities from literature	Building association databases from full-text articles

Integrated Data Strategies and Future Outlook

The emerging frontier in materials informatics leverages hybrid approaches that combine the reliability of curated data with the scale of text-mined knowledge. The most advanced synthesizability prediction frameworks, such as CSLLM, now integrate multiple specialized models—one for synthesizability classification trained on curated data, alongside separate models for predicting synthetic methods and precursors that can benefit from text-mined knowledge [16]. This integrated strategy addresses the multifaceted nature of synthesis prediction, where identifying a material as synthesizable represents only the first step toward experimental realization.

Future progress hinges on overcoming persistent challenges in both data paradigms. For curated data, limitations include incomplete coverage of chemical space and labor-intensive expansion processes [47]. Text mining faces hurdles in technical terminology processing and information veracity when applied to scientific literature [45]. The development of the Materials Ultimate Search Engine (MUSE) concept represents a visionary solution that would seamlessly integrate both data types, but requires community-wide standardization efforts and sustained investment in materials data infrastructure [42]. As these technical and institutional challenges are addressed, the complementary strengths of human-curated and text-mined approaches will continue to accelerate the discovery of novel functional materials through increasingly accurate synthesizability predictions.

Combating LLM Hallucinations with Domain-Specific Fine-Tuning

In the demanding field of computational materials science, particularly in predicting crystalline material synthesizability, the reliability of large language models is paramount. LLM hallucinations—fluent but factually incorrect or unsupported outputs—pose a significant barrier to trustworthy AI-assisted research [48] [49]. These hallucinations manifest as fabricated data, incorrect synthesizability predictions, or unsubstantiated precursor recommendations, potentially derailing experimental validation efforts. Domain-specific fine-tuning has emerged as a powerful methodology to combat these inaccuracies by aligning general-purpose LLMs with the precise terminology, relationships, and validation standards of specialized scientific domains. When applied to crystalline material synthesizability prediction, this approach demonstrates measurable improvements in factual consistency and predictive accuracy, creating more reliable research tools for materials scientists and drug development professionals working at the intersection of computational prediction and experimental synthesis.

Understanding LLM Hallucinations in Scientific Contexts

Defining and Classifying Hallucinations

Within scientific domains, LLM hallucinations present unique challenges due to the precise nature of technical information. Researchers categorize these inaccuracies into several distinct types [49]:

• Factual Hallucinations: Outputs containing scientifically inaccurate information, such as incorrect crystal formation energies or impossible synthetic pathways.
• Logical Hallucinations: Internally inconsistent reasoning, such as contradictory statements about thermodynamic stability.
• Extrinsic Hallucinations: Plausible-sounding information not grounded in the provided context or established scientific knowledge.

The mathematical foundation of these hallucinations stems from the probabilistic nature of LLMs, where the model may incorrectly assign higher probability to factually incorrect sequences than to accurate ones: Pθ(yhallucinated|x) > Pθ(ygrounded|x) [49]. In materials science applications, this miscalibration becomes particularly problematic when models generate synthesizability predictions or precursor recommendations that appear authoritative but lack experimental feasibility.

Domain-Specific Fine-Tuning: Methodologies and Approaches

Technical Foundations of Fine-Tuning

Domain-specific fine-tuning adapts general-purpose LLMs to specialized scientific domains through targeted training on curated datasets. Several technical approaches have demonstrated efficacy in reducing hallucinations in materials science contexts:

Supervised Fine-Tuning (SFT) involves continuing training of pre-trained models on domain-specific datasets, typically using a reduced learning rate (e.g., 1e-6) to preserve general capabilities while incorporating specialized knowledge [50]. This approach has proven particularly effective for crystalline materials prediction, where fine-tuned LLMs achieve state-of-the-art accuracy by learning domain-specific patterns from structured materials data [5] [4].

Parameter-Efficient Fine-Tuning methods, including LoRA (Low-Rank Adaptation), implement selective updates to model parameters, minimizing catastrophic forgetting while incorporating domain knowledge. These approaches are especially valuable when working with limited specialized datasets, as they prevent overfitting and maintain baseline model capabilities [50].

Domain-Specific Fine-Tuning Workflow

The following diagram illustrates the systematic workflow for domain-specific fine-tuning to combat hallucinations in materials science applications:

Experimental Comparison: Fine-Tuning Approaches for Synthesizability Prediction

Performance Metrics Across Methodologies

Multiple research studies have quantitatively evaluated the effectiveness of domain-specific fine-tuning for crystalline material synthesizability prediction. The following table summarizes key performance metrics across different approaches:

Table 1: Performance Comparison of LLM Fine-Tuning Methods for Crystalline Material Synthesizability Prediction

Fine-Tuning Method	Base Model	Accuracy (%)	Precision	Recall	Domain-Specific Benchmark Performance
StructGPT (SFT) [4]	GPT-4o-mini	95.8	0.942	0.961	Outperforms graph-based PU-CGCNN model
PU-GPT-embedding [4]	text-embedding-3-large	97.3	0.958	0.971	Superior to StructGPT and traditional graph methods
CSLLM Framework [5]	Specialized LLMs	98.6	N/A	N/A	Significantly outperforms thermodynamic (74.1%) and kinetic (82.2%) methods
Fine-tuning with small learning rate [50]	Various LLMs	Comparable to larger rates	Minimal general capability degradation	Preserved domain performance	Optimal balance for specialized adaptation

Hallucination Reduction Efficacy

The impact of domain-specific fine-tuning on hallucination reduction has been quantitatively measured across multiple studies:

Table 2: Hallucination Reduction Efficacy of Various Techniques in Materials Science Applications

Mitigation Technique	Hallucination Reduction (%)	Application Context	Implementation Complexity
Preference Optimization with Hallucination-Focused Datasets [51]	96%	General LLM applications	High
Retrieval-Augmented Generation (RAG) [48] [51]	70%	Crystallography data retrieval	Medium
RLHF with Calibrated Uncertainty Rewards [48] [51]	60%	Synthesizability prediction	High
Semantically-Driven Fine-Tuning [51]	50%	Materials property prediction	Medium
Adaptive Fact-Verification Algorithms [51]	40%	Experimental validation	Medium-High
Cross-Model Consensus Mechanisms [51]	30%	Multi-model validation systems	Medium

Experimental Protocols and Methodologies

Dataset Construction and Preparation

The foundation of effective domain-specific fine-tuning lies in meticulous dataset construction. For crystalline material synthesizability prediction, the following protocol has demonstrated efficacy [5] [4]:

• Positive Example Curation: 70,120 synthesizable crystal structures are selected from the Inorganic Crystal Structure Database (ICSD), filtered for structures containing ≤40 atoms and ≤7 different elements, with disordered structures excluded.
• Negative Example Generation: 80,000 non-synthesizable structures are identified from a pool of 1,401,562 theoretical structures using a pre-trained Positive-Unlabeled (PU) learning model with a CLscore threshold of <0.1.
• Text Representation: Crystal structures are converted to "material string" representations integrating space group information, lattice parameters (a, b, c, α, β, γ), and atomic site coordinates with Wyckoff positions.
• Data Balancing: The final dataset contains 150,120 structures representing seven crystal systems, with comprehensive elemental coverage (atomic numbers 1-94, excluding 85 and 87).

Fine-Tuning Implementation Protocol

The technical implementation of domain-specific fine-tuning follows a structured methodology [4] [50]:

• Model Selection: Base models (e.g., GPT-4o-mini, text-embedding-3-large) are selected based on architecture suitability and computational constraints.
• Learning Rate Configuration: A reduced learning rate (1e-6 to 5e-6) is employed to balance domain adaptation with general capability preservation.
• Tokenization Strategy: Domain-specific vocabulary is incorporated, with material representations tokenized to preserve structural information.
• Training Regimen: Models are trained for 3-5 epochs with batch sizes adapted to dataset size and computational resources.
• Validation Framework: Performance is evaluated using hold-out test sets with α-estimation for precision and false positive rate calculations in PU learning scenarios.

Hallucination Detection and Correction Framework

The HalluClean framework provides a systematic approach to identifying and addressing hallucinations in domain-specific LLM outputs [52]:

Successful implementation of domain-specific fine-tuning for hallucination reduction requires carefully selected computational resources and datasets:

Table 3: Essential Research Reagents for LLM Fine-Tuning in Materials Science

Resource/Reagent	Function	Access Method	Domain Relevance
ICSD (Inorganic Crystal Structure Database) [5] [4]	Provides experimentally verified crystal structures for positive examples	Commercial license	Ground truth for synthesizable materials
Materials Project Database [4]	Source of hypothetical structures for negative examples	Public API	Comprehensive coverage of calculated materials
Robocrystallographer [4]	Converts CIF files to text descriptions for LLM processing	Open-source Python package	Bridges structural data and natural language
PU Learning Models [5] [4]	Identifies non-synthesizable structures from unlabeled data	Custom implementation	Enables realistic negative example generation
Text-embedding-3-large [4]	Generates numerical representations from text descriptions	OpenAI API	Creates structured inputs for classifier models
HalluClean Framework [52]	Detects and corrects hallucinations in model outputs	Open-source implementation	Post-hoc verification of model predictions

Domain-specific fine-tuning represents a paradigm shift in combating LLM hallucinations for crystalline material synthesizability prediction. By systematically aligning general-purpose language models with the precise requirements of materials science, researchers can achieve unprecedented accuracy while maintaining factual integrity. The experimental evidence demonstrates that properly implemented fine-tuning strategies can reduce hallucination rates by up to 96% while achieving synthesizability prediction accuracy exceeding 98%, significantly outperforming traditional thermodynamic and kinetic stability assessments. As these methodologies continue to mature, they promise to accelerate the discovery and synthesis of novel materials by providing researchers with increasingly reliable AI assistants grounded in experimental feasibility and scientific rigor. The integration of structured reasoning frameworks, comprehensive domain datasets, and targeted verification protocols establishes a new standard for trustworthy AI in scientific applications, particularly in the critical domain of crystalline material synthesizability prediction where accuracy directly impacts experimental validation and resource allocation.

Predicting the synthesizability of crystalline materials is a critical step in accelerating the discovery of new functional materials for technologies ranging from pharmaceuticals to renewable energy. The traditional computational methods for assessing synthesizability have relied on principles of thermodynamic and kinetic stability, which, while physically grounded, can be computationally intensive and time-consuming. The emergence of machine learning (ML), and more recently, large language models (LLMs), offers a paradigm shift, promising to drastically reduce both the cost and time required for accurate predictions. This guide provides a objective comparison of these computational approaches, focusing on their relative accuracy, computational resource requirements, and associated costs. The evaluation is framed within the context of materials science and drug development, where rapid, reliable identification of synthesizable candidate materials can significantly compress research and development timelines. We present structured quantitative data, detailed experimental protocols, and clear visualizations to aid researchers and scientists in selecting the most efficient computational strategy for their specific needs.

Comparative Analysis of Predictive Performance

The performance of synthesizability prediction models is most critically judged by their accuracy, which measures the proportion of correct predictions (both synthesizable and non-synthesizable) across the entire dataset. Based on recent benchmarking studies, advanced computational models have demonstrated significant improvements over traditional methods.

The table below summarizes the key performance metrics and computational characteristics of the primary approaches:

Table 1: Performance and Cost Comparison of Synthesizability Prediction Methods

Prediction Method	Reported Accuracy	Relative Computational Cost	Primary Computational Resource	Typical Prediction Time per Structure
Traditional Stability Metrics
Thermodynamic (Energy Above Hull ≥0.1 eV/atom)	74.1% [5]	High	High-Performance Computing (HPC) Cluster	Hours to Days [6]
Kinetic (Phonon Frequency ≥ -0.1 THz)	82.2% [5]	Very High	High-Performance Computing (HPC) Cluster	Days [6]
Machine Learning (ML) Models
SynthNN (Composition-based)	Outperforms human experts (1.5x higher precision) [6]	Low	GPU-enabled Server	Minutes [6]
PU Learning Model (Structure-based)	87.9% [5]	Medium	GPU-enabled Server	Minutes [5]
Teacher-Student Dual NN	92.9% [5]	Medium	GPU-enabled Server	Minutes [5]
Large Language Models (LLMs)
Crystal Synthesis LLM (CSLLM)	98.6% [5]	Medium to High (for training) / Low (for inference)	GPU Cluster (Training) / GPU Server (Inference)	Seconds [5]

The data reveals a clear trajectory of increasing accuracy and speed from traditional methods to modern AI-driven approaches. The Crystal Synthesis Large Language Model (CSLLM) represents the state-of-the-art, achieving an accuracy of 98.6%, which substantially outperforms traditional thermodynamic (74.1%) and kinetic (82.2%) stability metrics [5]. Furthermore, LLMs and other ML models can generate predictions in seconds to minutes, a dramatic reduction from the hours or days required for density functional theory (DFT) calculations for energy or phonon spectra [5] [6]. This acceleration is a crucial enabler for high-throughput virtual screening of large material databases.

Beyond raw accuracy, the scope of prediction is an important differentiator. While traditional and many ML methods focus solely on a synthesizability score, the CSLLM framework demonstrates the capability to perform multiple interrelated tasks. Its specialized LLMs can predict not just synthesizability (98.6% accuracy), but also the appropriate synthetic method (91.0% classification accuracy) and identify suitable solid-state precursors (80.2% success rate) for common binary and ternary compounds [5]. This multi-task functionality provides a more comprehensive tool for experimental guidance.

Infrastructure, Deployment, and Cost Analysis

The computational efficiency of different prediction approaches is inextricably linked to the underlying hardware infrastructure and its associated costs. The shift from CPU-heavy traditional simulations to GPU-accelerated model inference has profound implications for both performance and budget.

Table 2: Computational Infrastructure and Cost Considerations (2025)

Infrastructure Component	Traditional HPC/DFT	AI/ML Model Inference	Notes & Cost Drivers
Primary Hardware	High-core count CPUs	GPUs (e.g., NVIDIA H100, A100)	AI ASICs are emerging alternatives to GPUs [53].
Cloud Compute Cost (Hourly)	Varies by CPU instance	~$2 - $15+ per GPU instance [54]	Cost depends on GPU type, memory, and provider [54].
Typical Workload Duration	Hours to days per structure [6]	Seconds to minutes per structure [5]	ML inference offers orders-of-magnitude speedup.
Total Cost of Workload	High (due to long runtimes)	Low (due to short runtimes)	"Cost per prediction" is a more useful metric than hourly rate [54].
Key Cost Optimization	Efficient parallelization	Model quantization, batching, use of spot instances [54]	Autoscaling can reduce idle resource costs for ML [55].

The financial investment in compute infrastructure is substantial and growing. The high-performance computing (HPC) market, which underpins these advanced research efforts, is forecast to grow by USD 23.45 billion between 2024 and 2029 [56]. A significant portion of this investment is directed towards GPU acceleration and AI-optimized hardware [56]. For context, the global data center processor market is projected to expand dramatically from nearly $150 billion in 2024 to over $370 billion by 2030, fueled by specialized hardware for AI workloads [53].

When deploying these models in the cloud, the "headline" hourly instance price is only one part of the cost equation. The total cost of inference is a more salient metric, which incorporates factors like throughput (predictions per second), latency, and GPU utilization [54]. Organizations can optimize these costs by employing techniques such as batching inference requests, using quantized models that require fewer resources, and leveraging a mix of on-demand, reserved, and spot instances from cloud providers [54]. Specialized GPU cloud platforms can sometimes offer lower latency and more predictable pricing than general-purpose hyperscalers for these specific tasks [54].

Experimental Protocols for Method Validation

To ensure the reliability and fair comparison of the different synthesizability prediction methods, rigorous experimental protocols are essential. The following sections outline the standard methodologies for training, validating, and benchmarking the leading approaches.

Protocol for Traditional Stability Calculations

• Structure Relaxation: Use density functional theory (DFT) with a standardized functional (e.g., PBE) and basis set to fully relax the candidate crystal structure, including lattice parameters and atomic positions, to its ground state [6].
• Energy Above Hull (Thermodynamic) Calculation:
  • Compute the formation energy of the candidate material.
  • Using a reference database (e.g., the Materials Project), construct the convex hull of formation energies for all other phases in the same chemical space.
  • The energy above hull is defined as the energy difference between the candidate and the convex hull at its composition. A value ≥0.1 eV/atom is often used as a stability threshold [5].
• Phonon Spectrum (Kinetic) Calculation:
  • Using the relaxed structure, compute the second-order force constants using the finite displacement method.
  • Calculate the phonon dispersion and density of states.
  • Analyze the spectrum for imaginary frequencies (soft modes). A lowest frequency ≥ -0.1 THz is a typical metric for dynamic stability [5].

Protocol for ML/LLM Model Training and Inference

• Dataset Curation:
  • Positive Examples: Curate synthesizable crystal structures from experimental databases like the Inorganic Crystal Structure Database (ICSD). For example, a dataset may include 70,120 structures, filtered for order and compositional diversity [5].
  • Negative Examples: Generate non-synthesizable examples by screening theoretical databases (e.g., the Materials Project) with a pre-trained model to identify structures with a low synthesizability score (e.g., CLscore <0.1). A balanced dataset might include 80,000 such structures [5].
• Feature Representation:
  • For composition-based models (e.g., SynthNN), use learned atom embeddings or stoichiometric attributes [6].
  • For structure-based LLMs (e.g., CSLLM), convert the crystal structure into a simplified text representation ("material string") that includes space group, lattice parameters, and Wyckoff positions for efficient model processing [5].
• Model Training:
  • Employ a Positive-Unlabeled (PU) learning framework to account for the fact that unsynthesized materials are not definitively unsynthesizable.
  • Fine-tune a base LLM (e.g., LLaMA) on the curated dataset of material strings and their synthesizability labels [5].
• Model Inference:
  • For a new candidate material, convert its structural information into the predefined text representation.
  • The model outputs a synthesizability classification and, if multi-task, predictions for synthesis method and precursors [5].

The workflow below illustrates the experimental pathway from a candidate crystal structure to a synthesizability prediction, highlighting the key differences between traditional and AI-driven methods:

Essential Research Toolkit

The following table details key resources and tools essential for conducting research in computational synthesizability prediction.

Table 3: Essential Research Reagents and Computational Tools

Item Name	Type	Function in Research	Example/Note
High-Performance Computing (HPC) Cluster	Infrastructure	Runs computationally intensive DFT calculations for traditional stability metrics.	Essential for phonon spectrum calculations [5].
GPU Cloud Instances	Infrastructure	Provides scalable computing for training large AI models and performing high-throughput inference.	Hourly cost ~$2-$15; optimized for parallelism [54].
ICSD (Inorganic Crystal Structure Database)	Data	The primary source of confirmed synthesizable crystal structures for training and benchmarking models [5] [6].	Contains over 70,000 curated structures [5].
Materials Project Database	Data	Provides a vast repository of computed crystal structures and properties, used for generating negative examples and convex hull data [5].	Contains data for over 1.4 million structures [5].
CIF (Crystallographic Information File)	Data Format	Standard text file format for representing crystal structure information.	Contains detailed lattice, atomic coordinate, and symmetry data [5].
Material String	Data Format	A simplified text representation of a crystal structure, designed for efficient processing by LLMs. Includes space group, lattice parameters, and Wyckoff positions [5].	Used by the CSLLM framework to reduce redundancy [5].
Pre-trained Large Language Model (LLM)	Software	A foundational model (e.g., LLaMA) that is fine-tuned on crystallographic data to create specialized synthesizability predictors [5].	Base for models like CSLLM [5].
AutoDock / SwissADME	Software (Related Field)	In-silico screening tools in drug discovery, representative of the computational approaches being adopted in materials science [57].	Used for virtual screening and predicting drug-likeness [57].

The comparative analysis presented in this guide clearly demonstrates a significant trade-off between computational cost and efficiency in predicting crystalline material synthesizability. Traditional DFT-based methods, while providing valuable thermodynamic and kinetic insights, require high computational costs and long timeframes, making them less suitable for the rapid screening of large material databases [5] [6]. In contrast, AI-driven approaches, particularly modern LLMs like the CSLLM framework, achieve superior accuracy (up to 98.6%) and reduce prediction times from days to seconds, albeit with a substantial upfront cost for model training and GPU infrastructure [5] [54] [53].

For researchers and drug development professionals, the choice of method should be guided by the project's specific stage and goals. Traditional methods remain invaluable for deep, mechanistic studies of a limited number of promising candidates. However, for the initial high-throughput discovery phase, where the goal is to quickly identify viable synthesizable materials from thousands or millions of candidates, LLMs and other ML models offer a transformative improvement in efficiency. The integration of multi-task prediction—providing not just a synthesizability score but also guidance on synthesis methods and precursors—further enhances the practical value of these AI tools, bridging the gap between computational prediction and experimental realization [5].

The acceleration of materials discovery through computational methods has created a critical bottleneck: experimental validation. While high-throughput screening can generate millions of candidate materials with promising properties, researchers lack the resources to synthesize and test them all. This challenge has spurred the development of predictive models that assess which theoretically proposed materials are likely to be synthesizable. However, as these models grow increasingly sophisticated, a fundamental question emerges: how do we interpret their predictions to gain genuine scientific insight rather than treating them as black boxes? The field of crystalline materials research now faces the dual challenge of not only achieving prediction accuracy but also ensuring model interpretability that can guide experimental synthesis efforts.

The distinction between prediction and explanation represents a core consideration in this domain. Predictive models focus primarily on accuracy, estimating the likelihood of future outcomes based on historical data [58]. In contrast, explanatory models aim to uncover underlying relationships between variables, often sacrificing some predictive power for interpretability and causal understanding [59]. For materials scientists, this distinction is crucial: knowing that a material is predicted to be synthesizable is useful, but understanding why it is synthesizable provides actionable insights that can inform synthesis strategies and guide the discovery of entirely new material classes.

Comparative Analysis of Synthesizability Prediction Approaches

Performance Metrics Comparison

The evaluation of synthesizability prediction models requires multiple metrics to capture different aspects of performance, from overall accuracy to clinical utility.

Table 1: Performance Metrics for Synthesizability Prediction Models

Metric	Definition	Interpretation	Ideal Value
Accuracy	Percentage of correctly classified instances	Overall correctness of predictions	Closer to 100%
Brier Score	Mean squared difference between predicted probabilities and actual outcomes	Overall model performance considering calibration	Closer to 0
C-statistic (AUC)	Area under Receiver Operating Characteristic curve	Discriminative ability to separate synthesizable/non-synthesizable	Closer to 1
Net Benefit	Weighted measure of true positives minus false positives	Clinical utility considering decision consequences	Higher than "all" or "none" strategies

Traditional statistical approaches for evaluating prediction models include the Brier score for overall model performance and the c-statistic (AUC) for discriminative ability [60]. More recently, decision-analytic measures such as net benefit and decision curve analysis have been proposed to evaluate the clinical utility of prediction models when used for decision-making [60] [61]. These are particularly relevant for synthesizability predictions, where researchers must decide which materials to prioritize for experimental validation.

Model Comparison for Crystalline Material Synthesizability

Recent advances in machine learning have produced several distinct approaches to predicting material synthesizability, each with different interpretability characteristics.

Table 2: Comparison of Synthesizability Prediction Models

Model	Accuracy	Interpretability Strength	Data Requirements	Key Limitations
CSLLM Framework [5]	98.6%	High (explicit precursor/method prediction)	150,120 crystal structures	Limited to 3D crystals with ≤40 atoms
SynthNN [6]	Not specified (7× higher precision than formation energy)	Medium (learns chemical principles)	Entire space of synthesized inorganic compositions	Structural information not utilized
PU Learning [11]	Varies by application	Medium (identifies likely synthesizable candidates)	Human-curated literature data	Limited by quality of text-mining

The Crystal Synthesis Large Language Models (CSLLM) framework represents a breakthrough in synthesizability prediction, achieving 98.6% accuracy by utilizing three specialized LLMs to predict synthesizability, synthetic methods, and suitable precursors respectively [5]. This approach significantly outperforms traditional screening methods based on thermodynamic stability (74.1% accuracy) and kinetic stability (82.2% accuracy). The framework's interpretability strength lies in its ability to not only classify materials as synthesizable but also provide specific, actionable guidance on how they might be synthesized.

In contrast, SynthNN adopts a different approach, leveraging the entire space of synthesized inorganic chemical compositions to predict synthesizability without requiring structural information [6]. Remarkably, this model learns fundamental chemical principles such as charge-balancing, chemical family relationships, and ionicity directly from the data distribution of previously synthesized materials. In head-to-head comparisons with human experts, SynthNN achieved 1.5× higher precision and completed the evaluation task five orders of magnitude faster than the best human expert.

Experimental Protocols and Methodologies

CSLLM Framework Implementation

The exceptional performance of the CSLLM framework stems from its sophisticated methodology and comprehensive dataset construction:

Dataset Construction: The model was trained on a balanced dataset containing 70,120 synthesizable crystal structures from the Inorganic Crystal Structure Database (ICSD) and 80,000 non-synthesizable structures identified from 1,401,562 theoretical structures using a pre-trained positive-unlabeled (PU) learning model [5]. To qualify for inclusion, structures could contain no more than 40 atoms and seven different elements, with disordered structures explicitly excluded.

Text Representation Innovation: A key innovation enabling the application of LLMs to crystal structures was the development of "material string" representation. This text format integrates essential crystal information—space group, lattice parameters, atomic species, Wyckoff positions, and fractional coordinates—in a compact, reversible format that eliminates redundancies present in CIF or POSCAR formats [5].

Model Architecture and Training: The framework employs three specialized LLMs fine-tuned on crystal structure data. The training process involved domain-focused fine-tuning to align the broad linguistic capabilities of foundation models with material-specific features critical to synthesizability, thereby refining attention mechanisms and reducing hallucinations [5].

CSLLM Framework Workflow: The process begins with comprehensive data collection from experimental and theoretical sources, transforms crystal structures into specialized text representations, fine-tunes LLMs on this data, and generates actionable synthesis predictions.

Positive-Unlabeled Learning Methodology

The application of positive-unlabeled (PU) learning to synthesizability prediction addresses a fundamental challenge in materials informatics: the absence of confirmed negative examples (definitively non-synthesizable materials) in literature data.

Data Processing: In the solid-state synthesis study by Chung et al., researchers manually curated synthesis information for 4,103 ternary oxides from literature, classifying them as solid-state synthesized (3,017 entries), non-solid-state synthesized (595 entries), or undetermined (491 entries) [11]. This human-curated dataset provided high-quality training data that enabled more accurate prediction of solid-state synthesizability compared to purely text-mined approaches.

Model Implementation: The PU learning approach treats un synthesized materials as "unlabeled" rather than definitively negative, probabilistically reweighting these examples according to their likelihood of being synthesizable [6] [11]. This semi-supervised approach acknowledges that materials absent from databases may be synthesizable but simply not yet discovered or reported.

Validation Framework: Model performance was evaluated through careful comparison with traditional thermodynamic metrics like energy above the convex hull (Ehull), revealing that thermodynamic stability alone is insufficient to predict synthesizability [11]. The integration of synthesis conditions and precursor information further enhanced predictive accuracy and interpretability.

Implementing interpretable synthesizability predictions requires specialized data resources and computational tools. The table below details key components of the research infrastructure supporting this field.

Table 3: Research Reagent Solutions for Synthesizability Predictions

Resource	Type	Primary Function	Key Features
ICSD [5] [11]	Database	Source of confirmed synthesizable structures	70,120+ curated crystal structures; experimental validation
Materials Project [5] [11]	Database	Repository of theoretical structures	1.4M+ calculated material structures; thermodynamic properties
Material String [5]	Data Representation	Text-based crystal structure encoding	Compact format; LLM-compatible; preserves symmetry information
PU Learning Algorithms [6] [11]	Computational Method	Learning from positive and unlabeled examples	Handles lack of negative data; probabilistic weighting
Decision Curve Analysis [60] [61]	Evaluation Framework	Assessing clinical utility of predictions	Incorporates consequence of decisions; threshold probabilities

These resources collectively enable the development and interpretation of predictive models that bridge computational materials design and experimental synthesis. The integration of multiple data sources—from carefully curated experimental databases to high-throughput computational repositories—provides the foundation for robust model training and validation.

Interpretation Frameworks: From Predictions to Scientific Insight

Navigating the Prediction-Explanation Spectrum

The fundamental distinction between predictive and explanatory modeling frameworks has profound implications for how researchers interpret synthesizability predictions:

Predictive modeling prioritizes accuracy metrics like root mean square error (RMSE) and focuses on forecasting outcomes based on historical patterns [59]. For synthesizability predictions, this approach can identify promising candidate materials but may offer limited insights into the underlying factors driving synthesizability.

Explanatory modeling emphasizes understanding variable relationships through inferential statistics like coefficient estimation and significance testing [59]. While potentially less accurate for prediction, this approach can reveal fundamental chemical principles that govern synthesizability.

The CSLLM framework occupies a middle ground, achieving high predictive accuracy while providing interpretable outputs through its specialized architecture. By separately predicting synthesizability, synthetic methods, and precursors, the model offers researchers multiple avenues for understanding its decision-making process [5].

Causal Inference Challenges

A critical limitation in interpreting predictive models involves distinguishing correlation from causation—a challenge particularly relevant for synthesizability predictions where multiple confounding factors may influence both input features and outcomes.

As illustrated in a subscription retention example, predictive models like XGBoost can identify robust correlations that nevertheless reflect reverse causality or unobserved confounding [62]. For instance, a model might find that materials with certain structural features are less likely to be synthesizable, when in reality these features correlate with research focus rather than intrinsic synthesizability.

Techniques like SHAP (SHapley Additive exPlanations) can make model decision processes more transparent by quantifying feature importance, but the Microsoft research team cautions that "making correlations transparent does not make them causal" [62]. For synthesizability predictions, this implies that while interpretation tools can highlight which structural descriptors most influence model predictions, establishing causal relationships requires additional experimental validation or carefully designed causal inference approaches.

Model Interpretation Decision Framework: Researchers must first determine whether their primary goal is prediction accuracy or scientific insight, then select appropriate interpretation methods, and ultimately validate interpretations through experimental synthesis.

The evolving landscape of synthesizability prediction demonstrates a clear trajectory toward models that balance predictive accuracy with scientific interpretability. The CSLLM framework's 98.6% accuracy represents a significant advancement over traditional thermodynamic and kinetic stability measures, while its specialized architecture provides researchers with specific, actionable insights into synthesis methods and precursor selection [5].

The integration of multiple evaluation approaches—from traditional discrimination and calibration metrics to decision-analytic measures like net benefit—provides a more comprehensive framework for assessing model utility in real-world research settings [60] [61]. As these interpretable models continue to evolve, they offer the promise of not only identifying synthesizable materials but also revealing fundamental principles of materials synthesis that can guide the discovery of entirely new material classes.

For researchers navigating this landscape, the critical consideration remains aligning model selection with research objectives: predictive models for identifying candidate materials, and explanatory approaches for understanding synthesis mechanisms. The most valuable frameworks, like CSLLM, integrate both capabilities—leveraging the pattern recognition power of advanced machine learning while maintaining interpretability that provides genuine scientific insight.

Benchmarking Model Accuracy: From Computational Metrics to Lab Validation

A Guide to Metrics for Evaluating Crystalline Material Synthesizability Predictions

For researchers navigating the complex challenge of predicting crystalline material synthesizability, selecting the right evaluation metric is not a mere formality—it is a critical decision that shapes model development and interpretation. This guide provides a head-to-head comparison of accuracy, precision, and recall, contextualized with experimental data from contemporary materials science research, to empower scientists in making informed choices.

---

In machine learning classification, a model's performance is fundamentally broken down using a confusion matrix, which categorizes predictions into True Positives (TP), True Negatives (TN), False Positives (FP), and False Negatives (FN) [63] [64]. Accuracy, precision, and recall are all derived from these core components but answer distinctly different questions about model behavior [65].

The table below summarizes the key characteristics, formulas, and primary use cases for each metric.

Metric	What It Measures	Formula	Ideal Use Case & Context
Accuracy	Overall correctness of the model [66] [63].	(TP + TN) / (TP + TN + FP + FN) [63] [65]	Balanced class distribution; cost of FP and FN is similar [64].
Precision	Reliability of positive predictions; how often a "positive" is correct [63] [64].	TP / (TP + FP) [66] [65]	False positives are costly (e.g., falsely claiming a material is synthesizable, wasting experimental resources) [66] [63].
Recall	Completeness of positive detection; ability to find all actual positives [63] [65].	TP / (TP + FN) [66] [65]	False negatives are costly (e.g., missing a truly synthesizable material, overlooking a promising candidate) [63] [65].

A critical limitation of accuracy is its susceptibility to misinterpretation under class imbalance, a common scenario in materials discovery where non-synthesizable candidates may vastly outnumber synthesizable ones. A model that simply labels all structures as "non-synthesizable" would achieve high accuracy while being practically useless for discovery, a phenomenon known as the accuracy paradox [64]. Precision and recall offer a more nuanced view by focusing specifically on the model's performance regarding the positive class of interest.

│ The Precision-Recall Trade-off and the F1 Score

In practice, it is often challenging to achieve both high precision and high recall simultaneously. This inherent tension is known as the precision-recall trade-off [63]. Adjusting a model's classification threshold can tune this balance: a higher threshold makes the model more conservative in making positive predictions, typically increasing precision but lowering recall; a lower threshold does the opposite, increasing recall but lowering precision [63].

The F1 score is a single metric that balances these two competing concerns. It is the harmonic mean of precision and recall and is particularly valuable when you need a single measure for model comparison and when both false positives and false negatives are important to avoid [67] [63] [65]. The formula for the F1 score is:

F1 Score = 2 × (Precision × Recall) / (Precision + Recall) [67] [66]

The following diagram illustrates the logical relationship between the confusion matrix, the core metrics, and the balancing function of the F1 score.

│ Experimental Protocols in Materials Synthesizability Prediction

Recent research has demonstrated the power of Large Language Models (LLMs) in predicting material properties and synthesizability. The experiments below showcase how evaluation metrics are applied in practice to validate such models.

The CSLLM Framework for Synthesizability Prediction

The Crystal Synthesis Large Language Model (CSLLM) framework was developed to accurately predict the synthesizability of 3D crystal structures, the likely synthetic method, and suitable precursors [5].

• Model Architecture & Input: The framework employs three specialized LLMs. The primary "Synthesizability LLM" was fine-tuned on a balanced dataset of 70,120 synthesizable structures from the Inorganic Crystal Structure Database (ICSD) and 80,000 non-synthesizable structures identified from a pool of over 1.4 million theoretical structures using a positive-unlabeled (PU) learning model [5]. The input to the model is a custom "material string"—a concise text representation of the crystal structure that includes space group, lattice parameters, and atomic coordinates with their Wyckoff positions [5].
• Evaluation & Results: The Synthesizability LLM was evaluated on a held-out test set. It achieved a remarkable 98.6% accuracy, significantly outperforming traditional screening methods based on thermodynamic stability (energy above hull, 74.1% accuracy) and kinetic stability (phonon spectrum, 82.2% accuracy) [5]. This high accuracy on a balanced test set indicates a model that is both highly correct overall and robust.

LLM-Prop for Crystal Property Prediction

The LLM-Prop model leverages the general-purpose learning capabilities of LLMs to predict various properties of crystals from their text descriptions [35].

• Model Architecture & Input: This approach uses the encoder part of a pre-trained T5 model (a Transformer-based architecture). The input is a textual description of the crystal structure. Key preprocessing steps included removing stopwords and replacing specific numerical values like bond distances and angles with special tokens ([NUM], [ANG]) to compress the sequence and allow the model to process longer contextual information [35].
• Evaluation & Results: LLM-Prop was benchmarked against state-of-the-art Graph Neural Networks (GNNs). It demonstrated superior performance on several regression tasks, outperforming the best GNN-based model by approximately 8% on predicting band gap and 65% on predicting unit cell volume, as measured by the relative reduction in Mean Absolute Error (MAE) [35]. This shows the model's high precision in predicting continuous numerical properties critical for materials design.

│ The Scientist's Toolkit: Essential Research Reagents

The following table details key computational tools and resources used in the featured experiments for crystalline materials research.

Tool/Resource	Function in Research
Large Language Models (LLMs) e.g., T5, LLaMA	Backbone architecture for understanding and processing textual or structured representations of crystal data, enabling property prediction and synthesizability classification [35] [5].
Text Representation (e.g., Material String, Processed Text)	Converts complex 3D crystal structure information into a standardized, machine-readable text format that serves as input for fine-tuned LLMs, ensuring efficient information capture [35] [5].
Inorganic Crystal Structure Database (ICSD)	A comprehensive database of experimentally reported crystal structures, serving as the primary source of confirmed "synthesizable" (positive) data points for training and benchmarking models [5].
Positive-Unlabeled (PU) Learning	A machine learning technique used to identify high-confidence "non-synthesizable" (negative) examples from large databases of theoretical structures, which is crucial for creating balanced training datasets [5] [68].
Graph Neural Networks (GNNs) e.g., ALIGNN, CGCNN	Established baseline models that represent crystals as graphs of atomic interactions; used as performance benchmarks for new approaches like LLM-based models [35] [69].

│ Choosing the Right Metric for Synthesizability Prediction

The choice of a primary evaluation metric should be guided by the specific research goal and the consequences of model errors.

• Prioritize Recall when the primary risk is missing a promising candidate. In early-stage discovery, where the cost of a false negative (overlooking a synthesizable material) is high, a high-recall model ensures comprehensive coverage, even if it means a higher rate of false alarms for experimental validation [65].
• Prioritize Precision when experimental resources are limited and costly. A high-precision model ensures that the candidates it identifies as synthesizable are highly likely to be correct, minimizing wasted effort on false leads [63].
• Use the F1 Score as a balanced metric for overall model comparison, especially when you need to ensure that neither precision nor recall is catastrophically low and when the cost of both false positives and false negatives is significant [67].
• Treat Accuracy with caution. It can be a useful high-level summary for a balanced dataset, but it should not be the sole metric for decision-making in the typically imbalanced context of materials discovery [65] [64].

In conclusion, the "best" metric is dictated by your research objective. By understanding the distinct role of accuracy, precision, and recall, and by leveraging modern LLM-based approaches, researchers can more effectively develop and deploy models that accelerate the discovery of novel crystalline materials.

In the accelerated discovery of new functional materials, a significant bottleneck persists: bridging the gap between computationally designed crystal structures and those that can be successfully synthesized in the laboratory. The journey from theoretical prediction to experimental realization hinges on accurately assessing crystallographic synthesizability—the likelihood that a proposed material can be experimentally realized. Traditional approaches have relied heavily on thermodynamic and kinetic stability metrics, such as formation energy and energy above the convex hull (Ehull) calculated via density functional theory (DFT), or the absence of imaginary phonon frequencies to indicate kinetic stability [5] [10]. However, these physical proxies alone show limited correlation with actual synthesizability because they fail to capture the complex, multifaceted nature of synthetic chemistry, which involves precursor selection, reaction pathways, and experimental conditions [5] [6]. This discrepancy has driven the emergence of machine learning (ML) models that learn synthesizability patterns directly from existing materials databases, with the Crystal Synthesis Large Language Model (CSLLM) representing a particularly advanced implementation achieving unprecedented 98.6% accuracy [5].

CSLLM: Architectural Framework and Methodology

The Three-Component LLM Framework

The CSLLM framework addresses the synthesizability challenge through a specialized, multi-component architecture [5]:

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

This tripartite structure enables CSLLM to provide comprehensive synthesis guidance beyond a simple binary classification, directly addressing the practical needs of experimental researchers.

Data Curation and Representation

A critical innovation underpinning CSLLM's performance lies in its sophisticated data strategy [5]:

• Positive Samples: 70,120 experimentally verified synthesizable crystal structures from the Inorganic Crystal Structure Database (ICSD), filtered to include structures with ≤40 atoms and ≤7 distinct elements.
• Negative Samples: 80,000 non-synthesizable structures identified from 1,401,562 theoretical structures across multiple databases using a pre-trained Positive-Unlabeled (PU) learning model, selecting those with crystal-likeness score (CLscore) <0.1.
• Novel Text Representation: Development of "material string" representation that efficiently encodes crystal structure information (space group, lattice parameters, atomic species, Wyckoff positions) in a compact text format suitable for LLM processing, eliminating redundancies present in CIF or POSCAR files.

Table: CSLLM Training Dataset Composition

Category	Data Source	Selection Criteria	Number of Structures
Synthesizable (Positive)	Inorganic Crystal Structure Database (ICSD)	≤40 atoms, ≤7 elements, ordered structures	70,120
Non-synthesizable (Negative)	Materials Project, Computational Material Database, OQMD, JARVIS	CLscore <0.1 from PU learning model	80,000
Total Training Data	-	-	150,120

Performance Comparison: CSLLM Versus Alternative Approaches

Quantitative Accuracy Assessment

CSLLM's synthesizability prediction capability demonstrates significant improvements over both traditional physical metrics and contemporary ML approaches [5]:

Table: Synthesizability Prediction Performance Comparison

Method	Principle	Accuracy/Performance	Key Limitations
CSLLM (Synthesizability LLM)	Fine-tuned large language model on material strings	98.6% accuracy on test set	Requires structured crystal information
Thermodynamic Stability (Ehull)	Energy above convex hull (DFT)	74.1% accuracy (≥0.1 eV/atom threshold)	Misses many synthesizable metastable phases
Kinetic Stability (Phonons)	Absence of imaginary frequencies in phonon spectrum	82.2% accuracy (≥ -0.1 THz threshold)	Computationally expensive; excludes some synthesizable materials
SynthNN	Deep learning on chemical compositions only	7× higher precision than DFT formation energies	Lacks structural information; lower precision than CSLLM
CPUL Model	Contrastive Positive-Unlabeled Learning	93.95% accuracy on MP test set	Lower accuracy than CSLLM; longer training required

Generalization Capability Assessment

A critical test of CSLLM's robustness involved evaluating its performance on complex crystal structures with complexity substantially exceeding its training data [5]. On this challenging generalization test, CSLLM maintained a remarkable 97.9% accuracy, demonstrating its ability to extract fundamental synthesizability principles rather than merely memorizing training patterns. The model also achieved exceptional results on the auxiliary tasks, with the Method LLM exceeding 90% accuracy in classifying synthetic methods, and the Precursor LLM achieving 80.2% success rate in identifying appropriate solid-state synthesis precursors for binary and ternary compounds [5].

Experimental Protocols and Validation Methodologies

CSLLM Training and Validation Protocol

The experimental methodology for developing and validating CSLLM followed a rigorous, multi-stage process [5]:

• Dataset Construction: Curated 150,120 crystal structures with balanced synthesizable/non-synthesizable examples across seven crystal systems and 1-7 elements.
• Feature Engineering: Transformed all crystal structures into standardized "material string" representations incorporating space group, lattice parameters, atomic species, and Wyckoff positions.
• Model Fine-tuning: Employed three separate LLMs (based on LLaMA-3 architecture) specialized for synthesizability classification, method prediction, and precursor identification.
• Performance Validation: Conducted standard train-test split validation followed by generalization testing on complex structures outside the training distribution.
• Comparative Analysis: Benchmarked against traditional methods (Ehull, phonons) and prior ML approaches using consistent evaluation metrics.

CSLLM Experimental Workflow

Benchmarking Protocol for Alternative Methods

To ensure fair comparison, researchers employed consistent benchmarking methodologies [5] [6] [46]:

• Traditional Methods: Thermodynamic stability assessed using Ehull ≥0.1 eV/atom as synthesizability threshold; kinetic stability assessed using lowest phonon frequency ≥ -0.1 THz threshold.
• ML Baselines: SynthNN evaluated using composition-only inputs on identical test sets; CPUL model assessed using its CLscore threshold of 0.5 for synthesizability classification.
• Generalization Testing: All methods evaluated on structures with complexity exceeding training data, particularly those with large unit cells and higher elemental diversity.

Table: Key Research Reagents and Computational Tools

Tool/Resource	Type	Function in Synthesizability Research	Example Sources
ICSD Database	Experimental Database	Provides experimentally verified synthesizable structures for training positive examples	ICSD [5]
Materials Project	Computational Database	Source of theoretical structures and calculated properties for negative sample generation	materialsproject.org [5]
PU Learning Models	Algorithm	Identifies non-synthesizable structures from unlabeled data for negative sample creation	CLscore model [5]
Material String Representation	Data Format	Compact text encoding of crystal structure for efficient LLM processing	CSLLM Framework [5]
Fine-tuned LLMs	Model Architecture	Specialized language models adapted for crystallographic synthesizability tasks	LLaMA-3 based CSLLM [5]
DFT Calculations	Computational Method	Provides traditional stability metrics (Ehull, phonons) for benchmark comparisons	VASP, Quantum ESPRESSO [10]

Implications for Materials Discovery Research

The demonstrated capabilities of CSLLM have substantial practical implications for high-throughput materials discovery. In one application, researchers leveraged CSLLM to screen 105,321 theoretical structures, successfully identifying 45,632 as synthesizable candidates [5]. These predicted synthesizable materials subsequently had 23 key properties calculated using graph neural networks, enabling efficient prioritization for experimental investigation. This end-to-end pipeline represents a significant acceleration over traditional discovery workflows.

Furthermore, CSLLM's architecture has been integrated into broader materials discovery frameworks such as T2MAT (text-to-material), where it serves as the synthesizability validation module that assesses generated structures and recommends synthesis pathways [70]. This integration highlights how CSLLM functions as a critical component bridging theoretical design and experimental realization in automated materials discovery platforms.

CSLLM's 98.6% prediction accuracy, coupled with its demonstrated generalization capability on structurally complex crystals, establishes a new state-of-the-art in computational synthesizability assessment. By significantly outperforming both traditional physical metrics and previous ML approaches, CSLLM addresses a critical bottleneck in materials discovery—the reliable identification of theoretically predicted structures that can be experimentally realized. The framework's multi-component architecture provides comprehensive synthesis guidance that extends beyond binary classification to include method selection and precursor identification, offering practical utility for experimental researchers. As materials research increasingly leverages AI-driven generative design and high-throughput computational screening, accurate synthesizability predictors like CSLLM will play an indispensable role in ensuring that theoretically promising materials can successfully transition from computational prediction to experimental realization and ultimately practical application.

The accelerating use of computational models to predict synthesizable crystalline materials has created a pressing need to evaluate their real-world performance. While accuracy metrics on benchmark test sets provide an initial quality signal, the ultimate validation occurs not in silicon but in the laboratory. This guide objectively compares contemporary synthesizability prediction methods, with a particular emphasis on experimental performance data that bridge the gap between computational promise and practical utility. As models evolve from thermodynamic proxies to sophisticated machine learning systems, their value for materials discovery must be measured by their ability to guide the synthesis of novel compounds under realistic conditions.

Comparative Performance of Synthesizability Prediction Methods

The table below summarizes the reported performance of various synthesizability prediction approaches, highlighting their methodological foundations and key quantitative metrics.

Table 1: Comparative Performance of Synthesizability Prediction Methods

Method/Model	Type	Key Innovation	Reported Test Accuracy	Experimental Success Rate
CSLLM [5]	Large Language Model	Three specialized LLMs for synthesizability, method, and precursors	98.6% accuracy	Not specified
SynthNN [6]	Deep Learning (Composition-based)	Positive-unlabeled learning from known compositions	7x higher precision than formation energy	Outperformed human experts (1.5x higher precision)
CPUL [46]	Contrastive + PU Learning	Combines contrastive learning with PU learning for feature extraction	93.95% accuracy (MP test set)	88.89% true positive rate (Fe-containing materials)
FTCP-SC [10]	Deep Learning (Structure-based)	Fourier-transformed crystal properties representation	82.6% precision / 80.6% recall (ternary crystals)	88.6% true positive rate (materials post-2019)
Synthesizability-Guided Pipeline [3]	Ensemble (Composition + Structure)	Rank-average ensemble of composition and structure models	High AUPRC on held-out test set	7 out of 16 targets successfully synthesized (44%)
Human Expert [6]	N/A	Baseline for comparison	N/A	Lower precision than SynthNN

Experimental Validation Protocols

A critical analysis of experimental validation methodologies reveals the rigor behind reported success rates.

High-Throughput Experimental Synthesis

The most direct validation involves selecting computationally predicted candidates and attempting their synthesis. A landmark 2025 study established a robust protocol, screening ~4.4 million computational structures to identify highly synthesizable candidates [3]. The experimental process was remarkably efficient, completing synthesis and characterization for 16 targets within just three days using an automated solid-state laboratory platform [3]. The successful synthesis of 7 previously unreported structures provides compelling evidence for the predictive utility of the underlying ensemble model.

Temporal Validation Splits

An alternative to immediate laboratory validation involves assessing performance on materials discovered after a model's training period. One study trained their model exclusively on compounds from the Materials Project database uploaded before 2015, then tested on materials added in subsequent years [10]. The model achieved an 88.6% true positive rate on the post-2019 dataset, demonstrating its ability to generalize to novel, real-world discoveries beyond its original training data [10].

Composition-Specific Generalization

Testing model performance on specific elemental systems absent from training data validates chemical transferability. The CPUL model was validated against all iron-containing materials in the Materials Project database, achieving an 88.89% true positive rate despite limited knowledge of Fe interactions in the training data [46]. This demonstrates robust learning of general synthesizability principles rather than mere memorization of training examples.

The Experimental Workflow for Validating Synthesizability Predictions

The following diagram illustrates the complete experimental pathway from computational prediction to laboratory validation, as implemented in state-of-the-art research.

Diagram 1: From Prediction to Synthesis Validation. This workflow illustrates the experimental validation pipeline, from screening millions of candidates to laboratory synthesis of selected targets [3].

The Scientist's Toolkit: Essential Research Reagents & Materials

Experimental validation of synthesizability predictions relies on specialized materials, databases, and computational resources.

Table 2: Essential Research Reagents and Resources for Synthesizability Research

Resource	Type	Primary Function	Example Sources/Composition
Solid-State Precursors	Chemical Reagents	Provide elemental components for synthesis reactions	Metal oxides, carbonates, other inorganic salts [11]
ICSD [5] [6]	Data Resource	Provides confirmed synthesizable structures for model training	Inorganic Crystal Structure Database
Materials Project [46] [3]	Data Resource	Source of theoretical structures & properties for prediction	DFT-calculated material database
Synthesizability Models	Computational Tool	Predicts likelihood of successful laboratory synthesis	CSLLM, SynthNN, CPUL, FTCP-SC [5] [6] [46]
High-Throughput Lab Platform	Equipment	Enables rapid synthesis of multiple candidates	Automated solid-state synthesis systems [3]
X-ray Diffractometer	Characterization	Verifies crystal structure of synthesized products	Laboratory or synchrotron X-ray source [3]

Multi-Model Framework for Synthesis Prediction

Beyond binary synthesizability classification, advanced frameworks now integrate multiple specialized models to predict various aspects of the synthesis process, as illustrated below.

Diagram 2: Multi-Model Synthesis Framework. Advanced systems like CSLLM employ specialized models for different synthesis aspects, providing comprehensive guidance beyond simple synthesizability classification [5].

The transition from theoretical accuracy to experimental validation represents the critical path for synthesizability prediction models. While test-set performance provides a necessary foundation, the true measure of utility emerges from laboratory synthesis outcomes. Current state-of-the-art models have demonstrated promising capabilities, with experimental success rates of approximately 44% in controlled, high-throughput studies [3]. This performance, while impressive, highlights both the progress made and the substantial room for improvement. Future advances will likely emerge from richer integration of synthesis route prediction, precursor identification, and condition optimization, moving beyond binary synthesizability classification toward comprehensive synthesis planning. For researchers relying on these tools, prioritizing models with demonstrated experimental validation, robust protocol documentation, and proven generalization to novel chemical systems remains essential for successful materials discovery.

Comparative Analysis of Model Performance Across Different Material Classes

The accelerated discovery of novel functional materials is a cornerstone of technological advancement. While computational methods, particularly density functional theory (DFT) and machine learning (ML), have successfully identified millions of candidate materials with promising properties, a significant bottleneck remains: predicting which theoretically proposed crystals are synthetically accessible [5]. The inability to reliably forecast synthesizability leads to a substantial gap between computational design and experimental realization, hindering the entire materials development pipeline.

Traditionally, synthesizability has been proxied by metrics of thermodynamic stability, such as energy above the convex hull (Ehull), or kinetic stability, assessed through phonon spectrum analysis [5] [10]. However, these approaches are imperfect; numerous metastable structures are successfully synthesized, while many thermodynamically stable structures remain elusive [6]. This discrepancy underscores the complex, multifaceted nature of synthesis, which is influenced by precursor choice, reaction pathways, and experimental conditions [5].

This guide provides a objective comparison of modern computational models developed to predict the synthesizability of crystalline inorganic materials. Framed within a broader thesis on accuracy metrics for this field, we analyze the performance of various approaches, from deep learning on composition to large language models (LLMs) fine-tuned on crystal structures. We focus on quantitative performance across different material classes, detail the experimental protocols behind benchmark results, and provide resources to equip researchers with the necessary tools for informed model selection.

The field has seen rapid evolution, from composition-based models to sophisticated structure-aware LLMs. The table below summarizes the performance of key models as reported in the literature.

Table 1: Comparative performance of synthesizability prediction models.

Model Name	Input Type	Architecture	Reported Accuracy (%)	Reported Precision (%)	Key Distinguishing Feature
CSLLM [5] [71]	Crystal Structure	Fine-tuned Large Language Model	98.6	N/A	Predicts synthesizability, method, and precursors
SynthNN [6]	Chemical Composition	Deep Learning (Atom2Vec)	N/A	~7x higher than DFT	Composition-only; no structure required
FTCP-based Model [10]	Crystal Structure	Deep Learning (Fourier Transform)	82.6 (Precision/Recall)	82.6	Uses combined real and reciprocal space features
Crystal Image CNN [72]	Crystal Structure (3D Image)	Convolutional Neural Network	High (exact % not specified)	N/A	Learns from image-based representation of crystals
CLscore (PU Learning) [5]	Crystal Structure	Positive-Unlabeled Learning	87.9	N/A	Used to generate non-synthesizable training data

The performance of these models is frequently compared against traditional stability metrics. For instance, the CSLLM framework has been shown to significantly outperform thermodynamic (energy above hull ≥0.1 eV/atom) and kinetic (lowest phonon frequency ≥ -0.1 THz) stability methods, surpassing them by 106.1% and 44.5% in accuracy, respectively [5] [71]. Similarly, SynthNN demonstrates a seven-fold higher precision in identifying synthesizable materials compared to using DFT-calculated formation energies alone [6].

Detailed Experimental Protocols

A critical factor in comparing model performance is understanding the experimental design and datasets used for training and validation. This section details the methodologies behind several key models.

The CSLLM Framework

The Crystal Synthesis Large Language Model (CSLLM) represents a recent advancement by employing a trio of fine-tuned LLMs for synthesizability, synthesis method, and precursor prediction [5] [71].

• Dataset Curation: A balanced dataset of 150,120 crystal structures was constructed. Positive (synthesizable) examples consisted of 70,120 ordered crystal structures from the Inorganic Crystal Structure Database (ICSD), filtered for structures with ≤40 atoms and ≤7 elements. Negative (non-synthesizable) examples were 80,000 structures with the lowest "crystal-likeness" scores (CLscore <0.1) screened from over 1.4 million theoretical structures in databases like the Materials Project (MP) using a pre-trained Positive-Unlabeled (PU) learning model [5].
• Text Representation: To enable LLM processing, a concise "material string" representation was developed. This format includes space group, lattice parameters, and a reduced set of atomic coordinates with Wyckoff positions, efficiently encoding crystal symmetry and information [5].
• Model Training and Testing: Three separate LLMs were fine-tuned on this dataset. The synthesizability LLM was tested on held-out data and achieved its 98.6% accuracy. Its generalization was further validated on complex structures with large unit cells, where it maintained 97.9% accuracy. The method and precursor LLMs achieved accuracies of 91.0% and 80.2%, respectively [5].

SynthNN and Composition-Based Learning

SynthNN addresses the challenge of predicting synthesizability when the crystal structure is unknown, relying solely on chemical composition [6].

• Dataset and PU Learning: The model is trained on chemical formulas from the ICSD, which are treated as positive examples. A key challenge is the lack of confirmed negative examples. This is addressed by augmenting the dataset with a large number of artificially generated, unsynthesized formulas and using a Positive-Unlabeled (PU) learning approach. This method treats the unlabeled (artificial) examples as probabilistically weighted negatives, accounting for the possibility that some might be synthesizable but undiscovered [6].
• Feature Representation: Instead of using predefined chemical descriptors, SynthNN employs an atom2vec representation. This method learns an optimal embedding for each atom directly from the distribution of synthesized materials in the ICSD, allowing the model to infer chemical principles like charge-balancing and ionicity from data [6].
• Benchmarking: Model performance is evaluated against baseline methods like random guessing and charge-balancing. In a head-to-head discovery challenge, SynthNN outperformed 20 expert materials scientists, achieving 1.5x higher precision and completing the task orders of magnitude faster [6].

FTCP and Crystal Graph Representations

Other structure-aware models use different strategies to convert crystal structures into machine-learnable features.

• Fourier-Transformed Crystal Properties (FTCP): This representation encodes crystals by creating feature vectors in both real space and reciprocal space. Real-space features use one-hot encoding, while reciprocal-space features are generated via a discrete Fourier transform of elemental property vectors. This hybrid approach aims to capture crystal periodicity and convoluted elemental properties that may be missed by other representations [10].
• Model and Performance: A deep learning classifier trained on FTCP representations achieved an overall precision and recall of 82.6% and 80.6%, respectively, for predicting synthesizability of ternary crystals. When trained on pre-2015 MP data and tested on materials added post-2019, the model identified a set of promising, unexplored candidates with a high true positive rate [10].

The following diagram illustrates the general workflow for the comparative evaluation of these different model architectures, from data preparation to performance assessment.

The Scientist's Toolkit

To facilitate practical implementation and reproducibility, the following table catalogues essential computational reagents and datasets used in developing and benchmarking synthesizability models.

Table 2: Key research reagents and resources for synthesizability prediction research.

Resource Name	Type	Primary Function	Reference/URL
Inorganic Crystal Structure Database (ICSD)	Database	Source of experimentally synthesized crystal structures for positive training examples.	[5] [6] [10]
Materials Project (MP)	Database	Repository of DFT-calculated structures and properties; source of hypothetical candidates.	[5] [10] [73]
Open Quantum Materials Database (OQMD)	Database	Another large-scale database of DFT-computed structures, used for training and validation.	[5] [73]
JARVIS	Database & Tools	Integrated platform for DFT, machine learning, and materials data. Hosts the JARVIS-Leaderboard.	[74]
JARVIS-Leaderboard	Benchmarking Platform	Community-driven platform for benchmarking various materials design methods (AI, DFT, FF).	https://pages.nist.gov/jarvis_leaderboard/ [74]
Matbench	Benchmarking Platform	Features a suite of predefined tasks for benchmarking ML models on materials property prediction.	[74]
CrabNet	Model/Algorithm	Composition-based property prediction model using self-attention mechanisms.	[10]
CGCNN	Model/Algorithm	Crystal Graph Convolutional Neural Network for property prediction from crystal structures.	[10]
ALIGNN	Model/Algorithm	Atomistic Line Graph Neural Network for accurate property prediction.	[74]

The comparative analysis presented in this guide reveals a dynamic and rapidly evolving field. The shift from traditional stability metrics to data-driven models has yielded significant improvements in prediction accuracy. Key trends include the move from composition-based to structure-aware models and the recent, groundbreaking application of large language models, which currently set the state-of-the-art in terms of reported accuracy and functional breadth [5] [71].

However, the choice of model is not one-size-fits-all. Researchers must consider the specific constraints of their discovery pipeline. For high-throughput screening of novel compositions where structure is unknown, composition-based models like SynthNN are indispensable [6]. When crystal structures are available, FTCP-based models or graph neural networks offer robust performance [10]. For the most comprehensive prediction, including guidance on synthesis routes and precursors, the CSLLM framework presents a powerful, albeit potentially more computationally intensive, option [5].

The ongoing development of integrated benchmarking platforms like the JARVIS-Leaderboard is crucial for ensuring rigorous, transparent, and reproducible comparisons between existing and future models [74]. As these tools mature and datasets expand, the reliable prediction of material synthesizability will continue to strengthen, finally closing the loop between computational design and experimental synthesis to accelerate the discovery of next-generation materials.

Conclusion

The field of crystalline material synthesizability prediction is rapidly maturing, transitioning from reliance on imperfect thermodynamic proxies to sophisticated data-driven models that achieve remarkable accuracy. The emergence of LLM-based frameworks and advanced PU-learning techniques demonstrates a clear path forward, with models like CSLLM reporting up to 98.6% accuracy. However, the ultimate validation of any model lies in its successful guidance of experimental synthesis, as evidenced by pipelines that have led to the creation of novel compounds. Future progress hinges on developing standardized benchmarks, improving the quality and scale of training data, and enhancing model explainability to build trust within the scientific community. For biomedical and clinical research, these advances promise to accelerate the discovery of novel functional materials for drug delivery systems, biomedical implants, and diagnostic tools, ultimately bridging the critical gap between in-silico design and real-world application.

References

• 1. SynCoTrain: a dual classifier PU-learning framework for ... [https://pubs.rsc.org/en/content/articlehtml/2025/dd/d4dd00394b]
• 2. Machine learning-based discovery of vibrationally stable ... [https://www.nature.com/articles/s41524-022-00943-z]
• 3. A Synthesizability-Guided Pipeline for Materials Discovery [https://arxiv.org/html/2511.01790v1]
• 4. Explainable Synthesizability Prediction of Inorganic Crystal ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12051750/]
• 5. Accurate prediction of synthesizability and precursors of 3D ... [https://www.nature.com/articles/s41467-025-61778-y]
• 6. Predicting the synthesizability of crystalline inorganic ... [https://www.nature.com/articles/s41524-023-01114-4]
• 7. Group News | Hemley - University of Illinois Chicago [https://hemley.lab.uic.edu/news-photos/]
• 8. The theory/calculation behind energy-above-hull [https://matsci.org/t/the-theory-calculation-behind-energy-above-hull-how-to-calculate-for-quarternary-structures/59743]
• 9. Explainable Synthesizability Prediction of Inorganic Crystal ... [https://chemrxiv.org/engage/chemrxiv/article-details/67565b8e7be152b1d05ee375]
• 10. Predicting Synthesizability using Machine Learning on ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9996807/]
• 11. Solid-state synthesizability predictions using positive ... [https://pubs.rsc.org/en/content/articlehtml/2025/dd/d5dd00065c]
• 12. Synthesizability of materials stoichiometry using semi- ... [https://www.sciencedirect.com/science/article/pii/S2590238524002273]
• 13. Accessible, Realistic, and Fair Evaluation of Positive ... [https://arxiv.org/html/2509.24228v1]
• 14. Positive and Unlabeled Materials Machine Learning [https://towardsdatascience.com/positive-and-unlabeled-materials-machine-learning-8b216edea899/]
• 15. Neighborhood-Based Positive Unlabeled Learning Using ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11120015/]
• 16. Accurate prediction of synthesizability and precursors of 3D ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12264133/]
• 17. Materials synthesizability and stability prediction using a semi ... [https://pubs.rsc.org/en/content/articlehtml/2023/dd/d2dd00098a]
• 18. Perovskite synthesizability using graph neural networks [https://www.nature.com/articles/s41524-022-00757-z]
• 19. Synthesizability Prediction of Crystalline Structures with a ... [https://www.arxiv.org/abs/2510.19251]
• 20. Structure-Based Synthesizability Prediction of Crystals ... [https://pubmed.ncbi.nlm.nih.gov/33104335/]
• 21. Kolmogorov–Arnold graph neural networks for molecular ... [https://www.nature.com/articles/s42256-025-01087-7]
• 22. MatGNet: A Graph Neural Network for Crystal Property ... [https://www.sciencedirect.com/science/article/abs/pii/S2352492825005331]
• 23. can graph neural network learn to explain activity cliffs? [https://pubs.rsc.org/en/content/articlehtml/2025/dd/d5dd00012b]
• 24. Graph neural networks for an accurate and interpretable ... [https://www.nature.com/articles/s41524-021-00574-w]
• 25. AI trends in 2025: Graph Neural Networks [https://assemblyai.com/blog/ai-trends-graph-neural-networks]
• 26. Performance assessment of various graph neural network ... [https://pubs.rsc.org/en/content/articlehtml/2025/cp/d5cp01072a]
• 27. A performance booster for Graph Neural Networks on node ... [https://www.sciencedirect.com/science/article/abs/pii/S0893608025001637]
• 28. 基于大语言模型的3D晶体结构可合成性及前驱体精准预测 ... [https://www.ebiotrade.com/newsf/2025-7/20250716081453437.htm]
• 29. 晶型预测到底有多准？ - ByDrug [https://bydrug.pharmcube.com/news/detail/4b87c29d3a69578f6274a9019a0fda0d]
• 30. DRAG让LLM准确率飙升45.5%，问题越复杂能力越强 [https://www.aigcopen.com/content/omni-channel/39228.html]
• 31. 大语言模型在科学研究中的应用概述 [https://www.istis.sh.cn/cms/news/article/45/27509]
• 32. SurFF：跨金属间化合物晶体表面暴露与形貌的基础模型 [https://cloud.tencent.com/developer/article/2576697]
• 33. 大语言模型（LLM）推理性能优化以及推理框架、后端的评测 [https://github.com/ninehills/blog/issues/107]
• 34. Machine-Learning-Assisted Prediction of Synthesizable ... [https://arxiv.org/html/2505.09161v1]
• 35. LLM-Prop: predicting the properties of crystalline materials ... [https://www.nature.com/articles/s41524-025-01536-2]
• 36. Leveraging language representation for materials ... [https://www.nature.com/articles/s41524-024-01231-8]
• 37. FastCSP: Accelerated Molecular Crystal Structure ... [https://arxiv.org/html/2508.02641v1]
• 38. Lead-free perovskites for next-generation applications [https://pubs.rsc.org/en/content/articlehtml/2025/ma/d5ma00681c]
• 39. Question Answering models for information extraction from ... [https://www.nature.com/articles/s43246-025-00979-w]
• 40. Advancing Solid - State of Inorganic Materials Synthesis [https://scisimple.com/en/articles/2025-09-26-advancing-solid-state-synthesis-of-inorganic-materials--aked5e1]
• 41. Machine-Learning Rationalization and Prediction of Solid - State ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9407029/]
• 42. Data‐Driven Materials Science: Status, Challenges, and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6839624/]
• 43. What is the difference between curated and text-mined ... [https://support.disgenet.com/support/solutions/articles/202000087529-what-is-the-difference-between-curated-and-text-mined-data-]
• 44. A comprehensive and quantitative comparison of text-mining ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5831415/]
• 45. Opportunities and challenges of text mining in materials ... [https://www.sciencedirect.com/science/article/pii/S2589004221001231]
• 46. Crystal synthesizability prediction using contrastive positive ... [https://www.sciencedirect.com/science/article/abs/pii/S0010465524003886]
• 47. Data science and material informatics in physical ... [https://www.sciencedirect.com/science/article/pii/S2590048X23000936]
• 48. LLM Hallucinations in 2025: How to Understand and ... [https://www.lakera.ai/blog/guide-to-hallucinations-in-large-language-models]
• 49. Survey and analysis of hallucinations in large language ... [https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2025.1622292/full]
• 50. Revisiting Domain-Specific Fine-Tuning in LLMs [https://arxiv.org/html/2509.20758v1]
• 51. Cutting-Edge Methods to Reduce LLM Hallucinations 2025 [https://sparkco.ai/blog/cutting-edge-methods-to-reduce-llm-hallucinations-2025]
• 52. A Unified Framework to Combat Hallucinations in LLMs [https://arxiv.org/html/2511.08916v3]
• 53. The Global Market for High Performance Computing (HPC ... [https://www.futuremarketsinc.com/the-global-market-for-high-performance-computing-hpc-and-ai-accelerators-2025-2035/]
• 54. How much does GPU cloud computing really cost in 2025? [https://www.gmicloud.ai/blog/how-much-does-gpu-cloud-computing-really-cost-in-2025]
• 55. The State Of AI Costs In 2025 [https://www.cloudzero.com/state-of-ai-costs/]
• 56. High Performance Computing (HPC) Market Size 2025-2029 [https://www.technavio.com/report/high-performance-computing-hpc-market-industry-analysis]
• 57. Top 5 Drug Discovery Trends 2025 Driving Breakthroughs [https://www.pelagobio.com/cetsa-drug-discovery-resources/blog/drug-discovery-trends-2025/]
• 58. Predictive Modeling & Statistical Analysis: Leveraging Data ... [https://online.nyit.edu/blog/predictive-modeling-statistical-analysis]
• 59. Prediction vs. Explanation: What's Your Model's Goal? [https://metricgate.com/blogs/prediction-vs-explanation-models/]
• 60. Assessing the performance of prediction models: a framework ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3575184/]
• 61. A simple, step-by-step guide to interpreting decision curve ... [https://diagnprognres.biomedcentral.com/articles/10.1186/s41512-019-0064-7]
• 62. Be Careful When Interpreting Predictive Models in Search ... [https://medium.com/data-science/be-careful-when-interpreting-predictive-models-in-search-of-causal-insights-e68626e664b6]
• 63. What is Accuracy vs. Precision vs. Recall in Machine ... [https://www.ultralytics.com/blog/accuracy-precision-recall]
• 64. Accuracy vs. precision vs. recall in machine learning [https://www.evidentlyai.com/classification-metrics/accuracy-precision-recall]
• 65. Classification: Accuracy, recall, precision, and related metrics [https://developers.google.com/machine-learning/crash-course/classification/accuracy-precision-recall]
• 66. Accuracy vs Precision vs Recall Explained with Real-World ... [https://medium.com/@backlinking2025/accuracy-vs-precision-vs-recall-explained-with-real-world-examples-0b54e1daf8e6]
• 67. The F1 Score: A Comprehensive Guide for 2025 [https://www.shadecoder.com/topics/the-f1-score-a-comprehensive-guide-for-2025]
• 68. Explainable Synthesizability Prediction of Inorganic Crystal ... [https://www.osti.gov/biblio/2536774]
• 69. LLM-mambaformer: Integrating mamba and transformer for ... [https://www.sciencedirect.com/science/article/abs/pii/S2352492825005410]
• 70. T2MAT (text-to-material): A universal agent for generating ... [https://www.sciencedirect.com/science/article/abs/pii/S1000681825001699]
• 71. [2407.07016] Is Large Language Model All You Need to ... [https://arxiv.org/abs/2407.07016]
• 72. Predicting synthesizability of crystalline materials via deep ... [https://www.nature.com/articles/s43246-021-00219-x]
• 73. Moving closer to experimental level materials property ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9279333/]
• 74. JARVIS-Leaderboard: a large scale benchmark of ... [https://www.nature.com/articles/s41524-024-01259-w]