Atom2Vec for Synthesizability Prediction: A Deep Learning Framework Accelerating Materials and Drug Discovery

Abstract

This article explores the transformative role of Atom2Vec and related deep learning representations in predicting the synthesizability of chemical compounds and materials. Tailored for researchers and drug development professionals, we cover the foundational principles of converting chemical formulas into machine-readable vectors, detail the methodology behind models like SynthNN and DeepSA, and address key challenges such as data scarcity and model interpretability. The content provides a comparative analysis against traditional methods, highlighting significant performance gains and real-world validation. This guide serves as a comprehensive resource for integrating these AI-driven tools into computational screening workflows to enhance the efficiency and success rate of discovering synthesizable candidates.

From Atoms to Algorithms: Demystifying Atom2Vec and the Synthesizability Challenge

The discovery of new materials and drug candidates is fundamentally limited by a critical challenge: the synthesizability problem. This refers to the significant gap between computationally designed compounds and their actual synthetic accessibility in a laboratory. In materials science and drug discovery, high-throughput computational methods can generate billions of candidate structures with desirable properties, but the vast majority are either impossible to synthesize with current methodologies or would require prohibitively complex synthesis pathways. The core issue stems from the fact that while computational models excel at predicting desirable properties from structure, they often lack the chemical intelligence to assess whether a proposed structure can be realistically constructed from available precursors and synthetic protocols.

Traditionally, assessing synthesizability has relied on expert knowledge and heuristic rules such as charge-balancing for inorganic materials. However, these approaches show limited accuracy; for instance, charge-balancing correctly identifies only 37% of known synthesized inorganic materials, and even among typically ionic binary cesium compounds, only 23% are charge-balanced [1]. Similarly, in drug discovery, conventional synthesizability scores often assume near-infinite building block availability, which does not reflect the resource-constrained environment of actual research laboratories [2]. The synthesizability problem therefore represents a major bottleneck in accelerating the discovery and development of new materials and therapeutics, necessitating advanced computational approaches that can integrate synthetic feasibility directly into the design process.

atom2vec Representation for Chemical Formulas

The atom2vec representation framework provides a powerful approach for encoding chemical information directly from material compositions without requiring prior structural knowledge. This method treats chemical formulas as foundational elements for a machine learning model, leveraging the entire space of synthesized inorganic chemical compositions to learn an optimal representation [1] [3]. Unlike traditional featurization methods that rely on pre-defined elemental descriptors, atom2vec learns embedding vectors for each atom type directly from the distribution of previously synthesized materials present in large databases.

In this approach, each chemical formula is represented by a learned atom embedding matrix that is optimized alongside all other parameters of a neural network [1]. The dimensionality of this representation is treated as a hyperparameter determined prior to model training. The key advantage of this method is that it requires no assumptions about which factors influence synthesizability or what metrics might serve as proxies for synthesizability, such as charge balancing or thermodynamic stability [1]. Instead, the chemical principles of synthesizability—including charge-balancing, chemical family relationships, and ionicity—are learned directly from the data of experimentally realized materials [1].

This data-driven representation is particularly valuable for predicting the synthesizability of novel compositions because it can capture complex patterns and relationships that may not be evident through traditional chemical intuition or rule-based approaches. The atom2vec model has demonstrated an ability to identify synthesizable materials with 7× higher precision than density-functional theory (DFT)-calculated formation energies, which are commonly used as a stability proxy [1].

Case Study: SynthNN - A Deep Learning Synthesizability Model

Model Architecture and Training Approach

SynthNN (Synthesizability Neural Network) exemplifies the application of atom2vec representation to predict the synthesizability of crystalline inorganic materials. This deep learning classification model is trained on a comprehensive dataset of chemical formulas derived from the Inorganic Crystal Structure Database (ICSD), which contains a nearly complete history of all crystalline inorganic materials reported in scientific literature [1]. To address the challenge that unsuccessful syntheses are rarely reported, the training dataset is augmented with artificially generated unsynthesized materials.

The model employs a semi-supervised learning approach that treats unsynthesized materials as unlabeled data and probabilistically reweights them according to their likelihood of being synthesizable [1]. This places SynthNN within the category of Positive-Unlabeled (PU) learning algorithms, which are particularly suited to materials science applications where negative examples (definitively unsynthesizable materials) are not available. The ratio of artificially generated formulas to synthesized formulas used in training is a key model hyperparameter (N_synth) that requires careful optimization [1].

Performance Benchmarking

SynthNN's performance has been rigorously evaluated against both computational methods and human experts:

Table 1: Performance Comparison of Synthesizability Prediction Methods

Method	Precision	Key Limitations
SynthNN	7× higher than DFT-based formation energy	Requires training data
Charge-Balancing	37% of known synthesized materials	Inflexible constraint
DFT Formation Energy	Captures only 50% of synthesized materials	Fails to account for kinetic stabilization
Human Experts	1.5× lower precision than SynthNN	Specialized domains, time-intensive

In a head-to-head material discovery comparison against 20 expert materials scientists, SynthNN outperformed all experts, achieving 1.5× higher precision and completing the task five orders of magnitude faster than the best human expert [1]. This demonstrates not only the accuracy but also the remarkable efficiency of the approach for high-throughput materials discovery.

Application Notes & Protocols

Protocol 1: Implementing SynthNN for Crystalline Materials

Purpose: To predict the synthesizability of novel inorganic crystalline compositions using the SynthNN framework.

Materials and Data Requirements:

• Inorganic Crystal Structure Database (ICSD): Primary source of synthesized material compositions for training [1]
• Artificially generated compositions: Created through combinatorial methods or negative sampling
• Computational resources: GPU-accelerated deep learning environment
• Software dependencies: Python with deep learning frameworks (PyTorch/TensorFlow)

Step-by-Step Procedure:

• Data Preparation:

  • Extract known synthesized compositions from ICSD
  • Generate artificial negative examples using combinatorial approaches
  • Split data into training (70%), validation (15%), and test (15%) sets

• Model Configuration:

  • Implement atom2vec embedding layer with dimensionality 64-256
  • Design neural network architecture with fully connected layers
  • Set hyperparameters (learning rate, batch size, N_synth ratio)

• Training Procedure:

  • Initialize model with random weights
  • Optimize using Adam optimizer with binary cross-entropy loss
  • Apply Positive-Unlabeled learning weighting for unlabeled examples
  • Validate model performance after each epoch

• Evaluation:

  • Calculate precision, recall, and F1-score on test set
  • Compare against charge-balancing and DFT formation energy baselines
  • Assess model calibration and confidence estimates

Troubleshooting:

• Poor performance may indicate inadequate representation of certain chemical spaces in training data
• Overfitting can be addressed through regularization techniques or data augmentation
• Embedding dimensionality can be optimized through hyperparameter tuning

Protocol 2: In-House Synthesizability Scoring for Drug Discovery

Purpose: To develop a synthesizability score tailored to specific available building blocks in drug discovery research.

Materials and Data Requirements:

• In-house building block inventory: Typically 5,000-10,000 available compounds [2]
• Computer-Aided Synthesis Planning (CASP) tool: e.g., AiZynthFinder [2]
• Target molecule set: For validation and training

Step-by-Step Procedure:

• Synthesis Planning Setup:

  • Configure CASP tool with in-house building block inventory
  • Set appropriate reaction parameters and constraints
  • Run synthesis planning on representative molecule set

• Data Generation:

  • Collect results of synthesis planning attempts
  • Label molecules as synthesizable or non-synthesizable based on CASP outcomes
  • Extract molecular features for training

• Model Training:

  • Implement machine learning classifier (random forest, neural network)
  • Train on CASP-derived synthesizability labels
  • Validate model predictions against held-out CASP results

• Integration:

  • Incorporate synthesizability score into de novo molecular design workflow
  • Use as filter or multi-objective optimization parameter

Performance Expectations: When using only ~6,000 in-house building blocks versus 17.4 million commercial compounds, synthesis planning success rates decrease by only approximately 12%, though synthesis routes may be two steps longer on average [2]. This minimal performance reduction demonstrates the feasibility of resource-limited synthesizability prediction.

Visualization: Synthesizability Prediction Workflow

Research Reagent Solutions

Table 2: Essential Resources for Synthesizability Research

Resource	Function	Application Context
Inorganic Crystal Structure Database (ICSD)	Provides known synthesized compositions for training	Crystalline inorganic materials [1]
AiZynthFinder	Computer-Aided Synthesis Planning tool	Drug discovery, organic molecules [2]
ZINC Database	Commercial building block repository	General synthesizability assessment [2]
Composition Analyzer Featurizer (CAF)	Generates numerical features from chemical formulas	Solid-state materials research [3]
Structure Analyzer Featurizer (SAF)	Extracts structural features from crystal structures	Structure-property mapping [3]
Matminer	Open-source toolkit for materials data featurization	High-throughput materials screening [3]

Validation and Experimental Design

Positive-Unlabeled Learning Framework

A critical consideration in synthesizability prediction is the absence of definitive negative examples, as materials not yet synthesized may still be synthesizable with future methodologies. The Positive-Unlabeled (PU) learning framework addresses this challenge by treating unsynthesized materials as unlabeled rather than definitively negative [1]. In this approach:

• Positive examples are known synthesized materials from databases like ICSD
• Unlabeled examples are artificially generated compositions not present in databases
• The model learns to distinguish synthesizable patterns while accounting for potential false negatives in the unlabeled set

PU learning algorithms probabilistically reweight unlabeled examples according to their likelihood of being synthesizable, improving model robustness against incomplete labeling [1]. Performance evaluation in this framework typically emphasizes F1-score rather than traditional accuracy metrics, as the true negative rate cannot be definitively established [1].

Experimental Validation Protocols

Experimental validation is essential to confirm computational synthesizability predictions. The following protocol outlines a systematic approach:

Purpose: To experimentally verify the synthesizability of computationally predicted materials or molecules.

Materials:

• Predicted synthesizable compositions/molecules
• Required precursors and building blocks
• Standard laboratory equipment for synthesis (e.g., solvothermal reactors, Schlenk lines)
• Characterization equipment (XRD, NMR, LC-MS)

Procedure:

• Synthesis Planning: Use CASP tools to identify specific synthesis routes
• Precursor Preparation: Acquire or synthesize required starting materials
• Synthesis Attempt: Execute proposed synthesis under varied conditions
• Characterization: Analyze products to confirm identity and purity
• Iterative Optimization: Modify conditions based on initial results

Case Study Example: In a recent study of monoglyceride lipase (MGLL) inhibitors, researchers experimentally evaluated three de novo candidates using AI-suggested synthesis routes employing only in-house building blocks [2]. They found one candidate with evident activity, demonstrating the practical utility of synthesizability-informed molecular design.

Computational Tools and Implementation

Available Featurization Tools

Multiple featurization tools are available for generating numerical representations from chemical compositions and structures:

Table 3: Comparison of Featurization Tools for Materials Research

Tool	Feature Type	Number of Features	Primary Applications
MAGPIE	Compositional	115-145	Perovskite discovery, superconducting critical temperature [3]
JARVIS	Compositional/Structural	438	2D materials identification, thermodynamic properties [3]
atom2vec	Compositional	N/A	Synthesizability prediction, crystal system classification [1] [3]
mat2vec	Compositional	200	Property prediction, materials natural language processing [3]
CAF+SAF	Compositional/Structural	227 total	Explainable ML for solid-state structures [3]
CGCNN	Structural	N/A	Crystal graph convolutional networks [3]

Visualization: Positive-Unlabeled Learning Approach

The integration of atom2vec representations with synthesizability classification models like SynthNN represents a significant advancement in addressing the synthesizability problem. These approaches leverage the complete landscape of known synthesized materials to learn the complex chemical principles governing synthetic accessibility, outperforming both traditional computational methods and human experts in prediction precision.

Future developments in this field will likely focus on several key areas:

• Multi-modal learning that integrates compositional, structural, and synthetic procedure data
• Transfer learning approaches to adapt synthesizability predictions to specific laboratory constraints
• Active learning frameworks that iteratively improve models based on experimental feedback
• Explainable AI techniques to elucidate the chemical rationale behind synthesizability predictions

As these methodologies mature, they will increasingly bridge the gap between computational design and experimental realization, accelerating the discovery of novel materials and therapeutic compounds with optimized properties and guaranteed synthetic feasibility.

The quest for effective machine representations of atoms constitutes a foundational challenge in computational materials science. Distributed representations characterize an object by embedding it in a continuous vector space, positioning semantically similar objects in close proximity [4]. For atoms, this means that elements sharing chemical similarities will reside near one another in this learned vector space. The Atom2Vec algorithm, introduced by Zhou et al., represents a groundbreaking approach to deriving these representations through unsupervised learning from extensive databases of known compounds and materials [5]. The core hypothesis is analogous to the natural language processing domain: just as words can be understood by the company they keep in textual corpora, atoms can be characterized by their common chemical environments in crystalline structures [4]. By leveraging the growing repositories of materials data, Atom2Vec learns the fundamental properties of atoms autonomously, without human supervision or pre-defined feature sets, generating high-dimensional vector representations that encapsulate complex chemical relationships and periodic trends.

Quantitative Performance Analysis

Performance of Atom2Vec and Related Models

Table 1: Comparative performance of different atom representation methods on materials property prediction tasks.

Model	Input Data	Representation Dimensionality	Key Advantages	Limitations
Atom2Vec [5] [4]	Crystal structure databases	Varies (hyperparameter)	Discovers chemical similarities without prior knowledge	Limited to elements in training data
SkipAtom [4]	Crystal structure graphs using Voronoi decomposition	Not specified	Reflects chemo-structural environments; effective for compound representation	Requires structural data for training
Mat2Vec [4]	Scientific abstracts from materials literature	Not specified	Captures research context and trends	May reflect scientific interest rather than inherent chemistry
Element2Vec [6]	Wikipedia text using LLMs	Global and local embeddings	Incorporates rich textual knowledge; interpretable attributes	Limited by quality and coverage of source text
Random Vectors [4]	Random sampling from standard normal distribution	Arbitrary	Simple to generate	No semantic relationships
One-Hot Vectors [4]	Unique binary vectors per element	Number of element categories	Simple interpretation	No similarity information; high dimensionality

Synthesizability Prediction Performance

Table 2: Performance comparison of synthesizability prediction methods on inorganic crystalline materials.

Method	Basis of Prediction	Precision	Key Findings	Reference
SynthNN (with Atom2Vec)	Deep learning on known compositions	7× higher than formation energy	Outperformed all 20 human experts; 1.5× higher precision than best expert	[1]
Charge-Balancing	Net neutral ionic charge using common oxidation states	Low (23-37% of known compounds)	Inflexible; cannot account for different bonding environments	[1]
DFT Formation Energy	Thermodynamic stability relative to decomposition products	Captures only 50% of synthesized materials	Fails to account for kinetic stabilization	[1]
BLMM Crystal Transformer	Blank-filling language model	89.7% charge neutrality, 84.8% balanced electronegativity	4-8× higher than pseudo-random sampling	[7]

Fundamental Atom2Vec Methodology

Core Algorithm and Workflow

The Atom2Vec algorithm employs an unsupervised learning framework inspired by natural language processing techniques. The fundamental analogy translates words in sentences to atoms in crystal structures [4]. The methodology involves these key steps:

• Data Extraction: Gather crystal structures from comprehensive materials databases such as the Inorganic Crystal Structure Database (ICSD) [1].

• Environment Identification: For each atom in every crystal structure, identify its local chemical environment, typically defined by its nearest neighbors or coordination sphere.

• Co-occurrence Matrix Generation: Construct a matrix documenting how frequently atoms co-occur in similar chemical environments across all structures in the database [4].

• Dimensionality Reduction: Apply singular value decomposition (SVD) or neural network-based embedding techniques to the co-occurrence matrix to obtain lower-dimensional vector representations for each atom [4].

The resulting atom vectors capture complex chemical relationships, with atoms sharing similar properties or positions in the periodic table naturally clustering together in the vector space [5].

Implementation for Synthesizability Prediction

The application of Atom2Vec to synthesizability prediction involves specific adaptations and training strategies, as exemplified by the SynthNN model [1]:

• Positive-Unlabeled Learning: The model is trained on a dataset consisting of:

  • Positive examples: Chemical formulas extracted from the ICSD representing synthesized materials [1].
  • Artificially generated unsynthesized materials: Treated as unlabeled data and probabilistically reweighted according to their likelihood of being synthesizable [1].

• Representation Learning: Chemical formulas are represented using the learned atom embedding matrix, which is optimized alongside all other parameters of the neural network [1].

• Model Architecture: A deep neural network (SynthNN) is trained to classify materials as synthesizable or not based on their compositional representations, without requiring structural information [1].

This approach allows the model to learn the implicit "chemistry of synthesizability" directly from the distribution of experimentally realized materials, capturing complex factors beyond simple charge-balancing or thermodynamic stability [1].

Advanced Experimental Protocols

Protocol 1: Atom2Vec Model Training

Objective: Train Atom2Vec embeddings from a crystalline materials database.

Materials and Input Data:

• Crystal Structure Database: Inorganic Crystal Structure Database (ICSD) or Materials Project database [1].
• Computing Environment: Standard deep learning framework (e.g., TensorFlow, PyTorch).
• Preprocessing Tools: Voronoi decomposition algorithms for identifying atomic neighbors in crystal structures [4].

Procedure:

• Data Preparation:
  • Extract crystal structures from the chosen database.
  • For each structure, identify all atomic pairs within a specified cutoff distance or using Voronoi tessellation to determine nearest neighbors [4].

• Training Set Generation:

  • Create training pairs consisting of (target atom, context atom) for each atomic environment.
  • For each atom in every crystal structure, pair it with all its neighboring atoms identified in step 1.

• Model Configuration:

  • Set embedding dimension as a hyperparameter (typically 50-200 dimensions).
  • Initialize atom vectors randomly or using pre-trained values if available.

• Model Training:

  • Use Maximum Likelihood Estimation to maximize the average log probability: (1/|M|) Σ_{m∈M} Σ_{a∈A_m} Σ_{n∈N(a)} log p(n|a) where M is the set of materials, A_m is the set of atoms in material m, and N(a) are the neighbors of atom a [4].
  • Minimize cross-entropy loss between the one-hot vector representing the context atom and the normalized probabilities produced by the model.

• Validation:

  • Evaluate learned embeddings by examining clustering of chemically similar elements.
  • Test performance on downstream tasks such as formation energy prediction.

Protocol 2: Synthesizability Prediction with SynthNN

Objective: Predict synthesizability of inorganic chemical formulas using Atom2Vec representations.

Materials and Input Data:

• Positive Examples: 180,000+ synthesized inorganic crystalline materials from ICSD [1].
• Negative Examples: Artificially generated unsynthesized materials (treated as unlabeled data in PU learning framework).
• Atom2Vec Embeddings: Pre-trained atom vectors.

Procedure:

• Dataset Construction:
  • Compile chemical formulas of known synthesized materials from ICSD as positive examples.
  • Generate artificial negative examples through combinatorial composition generation or random sampling of chemical formulas not present in ICSD.
  • Apply positive-unlabeled learning techniques to account for potentially synthesizable materials among the unlabeled examples [1].

• Feature Representation:

  • Represent each chemical formula using Atom2Vec embeddings of constituent atoms.
  • Use pooling operations (e.g., sum, average, weighted average) to combine atomic vectors into fixed-dimensional compound representations [4].

• Model Architecture:

  • Implement a deep neural network classifier (SynthNN) with multiple fully connected layers.
  • Use ReLU or similar activation functions between layers.
  • Apply appropriate regularization techniques (dropout, L2 regularization) to prevent overfitting.

• Training Procedure:

  • Train the model to distinguish between synthesized and artificially generated materials.
  • Use class weighting or sampling techniques to handle dataset imbalance.
  • Employ early stopping based on validation performance.

• Model Evaluation:

  • Compare performance against baseline methods (charge-balancing, formation energy) [1].
  • Evaluate precision, recall, and F1-score on holdout test set.
  • Conduct comparative studies with human experts to benchmark practical utility [1].

Essential Research Reagents and Computational Tools

Table 3: Key resources for implementing Atom2Vec and synthesizability prediction models.

Resource Category	Specific Tools/Databases	Function and Application
Materials Databases	Inorganic Crystal Structure Database (ICSD), Materials Project, OQMD	Source of crystal structures for training Atom2Vec models [1] [7]
Representation Models	Atom2Vec, SkipAtom, Mat2Vec, Element2Vec	Provide atomic embeddings for materials informatics tasks [5] [4] [6]
Deep Learning Frameworks	TensorFlow, PyTorch, JAX	Implementation of neural networks for SynthNN and related models
Language Models	BLMM Crystal Transformer, MatSciBERT	Alternative approaches for materials representation and generation [7]
Property Prediction Models	CGCNN, MEGNet, ALIGNN, PotNet	Benchmark models for evaluating quality of learned representations [8]
Generative Models	CDVAE, DiffCSP, GNoME	For inverse design of materials using learned representations [8]

Future Directions and Advanced Applications

The integration of Atom2Vec with emerging deep learning architectures presents promising avenues for advancement. Transformer-based models like the Blank-filling Language Model for Materials (BLMM) have demonstrated exceptional capability in generating chemically valid compositions with high charge neutrality (89.7%) and balanced electronegativity (84.8%) [7]. These models effectively learn implicit "materials grammars" from composition data, enabling interpretable design and tinkering operations. For drug development professionals, these approaches facilitate rapid exploration of chemical space for novel inorganic compounds with potential pharmaceutical applications, such as contrast agents or diagnostic materials. The continuing development of inverse design pipelines using generative models trained on Atom2Vec representations will further accelerate the discovery of synthesizable materials with targeted properties [9]. As these models evolve, they increasingly capture complex chemical principles including charge-balancing, chemical family relationships, and ionicity, providing powerful tools for rational materials design across scientific disciplines [1].

Predicting whether a hypothetical chemical compound can be successfully synthesized remains a fundamental challenge in materials science and drug discovery. For decades, charge-balancing—ensuring a net neutral ionic charge based on elements' common oxidation states—has served as a primary heuristic for assessing synthesizability. However, evidence from large-scale materials databases reveals that this traditional metric fails to accurately predict synthetic accessibility. Analysis of the Inorganic Crystal Structure Database (ICSD) shows that only 37% of all known synthesized inorganic compounds are charge-balanced according to common oxidation states, with this figure dropping to a mere 23% for binary cesium compounds [1]. This discrepancy underscores a critical limitation: while chemically intuitive, charge-balancing operates as an excessively rigid filter that cannot account for the diverse bonding environments present in metallic alloys, covalent materials, or kinetically stabilized phases [1].

The development of atomistic representation learning methods, particularly atom2vec and its derivatives, has enabled more sophisticated, data-driven approaches to synthesizability prediction. These techniques learn distributed representations of atoms from existing materials databases, capturing complex chemical relationships that extend beyond simple charge-balancing considerations. By reformulating material discovery as a synthesizability classification task, models like SynthNN (Synthesizability Neural Network) demonstrate 7× higher precision than traditional formation energy calculations and outperform human experts by achieving 1.5× higher precision while completing screening tasks five orders of magnitude faster [1]. This Application Note examines the limitations of traditional synthesizability metrics and provides detailed protocols for implementing advanced machine learning approaches that leverage learned atomic representations.

Limitations of Traditional Synthesizability Metrics

Quantitative Comparison of Synthesizability Prediction Methods

Table 1: Performance comparison of synthesizability prediction approaches

Method	Key Principle	Precision	Recall	Key Limitations
Charge-Balancing	Net neutral ionic charge based on oxidation states	Low (exact values not reported)	N/A	Overly rigid; ignores diverse bonding environments; only 37% of known materials comply [1]
DFT Formation Energy	Thermodynamic stability relative to decomposition products	~4× lower than SynthNN [1]	~50% of synthesized materials [1]	Fails to account for kinetic stabilization; computationally intensive
SynthNN (atom2vec)	Data-driven classification from known materials	7× higher than DFT [1]	High (outperforms human experts) [1]	Requires sufficient training data; model interpretability challenges
FTCP Representation	Crystal structure representation in real/reciprocal space	82.6% [10]	80.6% [10]	Requires structural information; less effective for composition-only screening
SC Model	Synthesizability score from structural fingerprints	88.6% TPR for post-2019 materials [10]	9.81% precision for post-2019 materials [10]	Performance varies across chemical spaces

Why Charge-Balancing and Thermodynamic Metrics Fail

Traditional synthesizability assessment suffers from several fundamental limitations that machine learning approaches directly address:

• Oversimplified Chemical Intuition: Charge-balancing applies a one-size-fits-all approach that fails to capture material-specific bonding characteristics. The metric performs particularly poorly for materials with metallic bonding, complex covalent networks, or those stabilized through kinetic rather than thermodynamic pathways [1].

• Incomplete Stability Assessment: While DFT-calculated formation energy and energy above hull (Ehull) provide valuable thermodynamic insights, they fail to account for kinetic stabilization, synthetic pathway availability, and practical experimental constraints [10]. Studies indicate that formation energy calculations capture only approximately 50% of synthesized inorganic crystalline materials [1].

• Exclusion of Practical Considerations: Traditional metrics ignore crucial experimental factors including precursor availability, equipment requirements, earth abundance of starting materials, toxicity considerations, and human-perceived importance of the target material—all factors that significantly influence synthetic decisions [1] [10].

• Inability to Generalize: Rule-based approaches lack the flexibility to adapt to new chemical spaces or evolving synthetic methodologies, whereas data-driven models continuously improve as additional synthesized materials are reported [1].

Atomistic Representation Learning for Synthesizability

Foundational Concepts and Methodologies

Atomistic representation learning methods transform chemical elements into continuous vector embeddings that capture nuanced chemical relationships, mirroring successful natural language processing approaches where words with similar contexts have similar vector representations [4]. These learned representations form the foundation for modern synthesizability prediction models.

Table 2: Key atomic representation learning methods

Method	Training Data	Representation Dimensionality	Key Advantages
Atom2Vec	Co-occurrence matrix from materials database	Limited to number of atoms in matrix [4]	Captures elemental relationships from crystal structures
Mat2Vec	Scientific abstracts (text corpus) [4]	200 dimensions [3]	Leverages rich contextual information from literature
SkipAtom	Crystal structures with atomic connectivity graphs [4]	Configurable (typically 50-200 dimensions)	Explicitly models local chemical environments; unsupervised
Element2Vec	Wikipedia text pages [11]	Global and local embeddings	Combines holistic and attribute-specific information

The SynthNN Architecture and Workflow

The SynthNN model exemplifies the application of atomistic representations to synthesizability prediction. This deep learning framework leverages the entire space of synthesized inorganic chemical compositions through the following workflow:

Research Reagent Solutions: Essential Computational Tools

Table 3: Key software and data resources for synthesizability prediction

Resource	Type	Function	Access
Inorganic Crystal Structure Database (ICSD)	Structured database	Source of synthesized materials for training [1] [10]	Commercial
Materials Project API	Computational database	DFT-calculated properties; structural information [10]	Open
atom2vec	Algorithm	Unsupervised atomic representation learning [4]	Open
Matminer	Featurization toolkit	Compositional and structural descriptor generation [3]	Open
CAF/SAF	Feature generators	Composition Analyzer Featurizer (CAF) and Structure Analyzer Featurizer (SAF) for explainable ML [3]	Open
BLMM Crystal Transformer	Blank-filling language model	Generative design of inorganic materials [7]	Open

Experimental Protocols

Protocol 1: Implementing SynthNN for Synthesizability Classification

Purpose: To train a deep learning model for synthesizability prediction using atomistic representations and known materials data.

Materials and Data Sources:

• Inorganic Crystal Structure Database (ICSD) for synthesized materials [1]
• Artificially generated unsynthesized compositions for negative examples [1]
• Python deep learning framework (PyTorch/TensorFlow)
• Atom2Vec or SkipAtom pretrained embeddings [4]

Procedure:

• Data Preparation:

  • Extract chemical formulas of known synthesized inorganic crystalline materials from ICSD (approximately 20,000-200,000 compositions) [1].
  • Generate artificial negative examples through combinatorial composition generation or by sampling from hypothetical materials databases.
  • Apply positive-unlabeled (PU) learning techniques to account for potentially synthesizable materials among the unlabeled examples [1].

• Feature Generation:

  • Implement atom2vec embedding lookup for each element in the chemical formula.
  • Apply pooling operations (sum, average, or weighted pooling) to create fixed-length composition representations [4].
  • For comparative analysis, generate alternative features including Magpie descriptors, one-hot encodings, or mat2vec representations [3].

• Model Architecture:

  • Construct a deep neural network with 3-5 hidden layers (512-1024 neurons per layer) with ReLU activation functions.
  • Include dropout layers (rate=0.3-0.5) to prevent overfitting.
  • Implement batch normalization between layers for training stability.
  • Use sigmoid activation in the final layer for binary classification.

• Model Training:

  • Employ stratified k-fold cross-validation (k=5-10) to evaluate model performance.
  • Utilize Adam optimizer with learning rate 0.001-0.0001 and binary cross-entropy loss.
  • Implement early stopping with patience of 10-20 epochs based on validation loss.
  • Balance training batches to address class imbalance.

• Model Evaluation:

  • Calculate precision, recall, F1-score, and AUC-ROC metrics.
  • Compare performance against charge-balancing and DFT-based baselines.
  • Perform ablation studies to assess contribution of different embedding strategies.

Troubleshooting:

• For poor convergence: Adjust learning rate, try different embedding dimensions, or increase model capacity.
• For overfitting: Increase dropout rate, implement L2 regularization, or augment training data.
• For embedding instability: Use pretrained embeddings or increase embedding training data.

Protocol 2: Composition-Based Synthesizability Screening

Purpose: To rapidly screen novel chemical compositions for synthesizability potential using only composition information.

Materials and Data Sources:

• Pretrained SynthNN model or alternative synthesizability classifier
• Candidate compositions for screening (e.g., from generative models or combinatorial enumeration)
• BLMM Crystal Transformer for composition generation and validation [7]

Procedure:

• Composition Generation:

  • Generate candidate compositions using generative models (GAN, VAE, or transformer-based).
  • Apply charge-balancing as an initial filter (despite limitations) to reduce candidate space.
  • Utilize BLMM for composition generation with built-in charge neutrality and electronegativity constraints [7].

• Feature Extraction:

  • Parse chemical formulas into constituent elements and stoichiometric coefficients.
  • Convert elements to embeddings using pretrained atom2vec, SkipAtom, or similar.
  • Apply stoichiometry-weighted pooling to create fixed-length vectors.
  • Alternatively, use Magpie features or one-hot encodings for baseline comparisons.

• Synthesizability Prediction:

  • Apply trained classification model to generate synthesizability scores (0-1 scale).
  • Rank candidates by synthesizability score for prioritization.
  • Implement ensemble methods by combining predictions from multiple models.

• Validation:

  • Compare predictions with DFT-calculated formation energies where feasible.
  • For top candidates, perform structural prediction and further stability analysis.
  • Select highest-confidence candidates for experimental validation.

Troubleshooting:

• For chemically implausible high-scoring compositions: Incorporate additional constraints (electronegativity balance, radius ratio rules).
• For inconsistent predictions across similar compositions: Ensure embedding stability and consider ensemble methods.
• For memory limitations with large screening sets: Implement batch processing and efficient vector operations.

Protocol 3: Explainable Synthesizability Analysis

Purpose: To interpret synthesizability predictions and identify contributing chemical factors.

Materials and Data Sources:

• Trained synthesizability model
• Composition Analyzer Featurizer (CAF) and Structure Analyzer Featurizer (SAF) [3]
• Model interpretation libraries (SHAP, LIME)

Procedure:

• Feature Importance Analysis:

  • Apply SHAP (SHapley Additive exPlanations) to quantify feature contributions.
  • Identify which elements and stoichiometric ratios most influence predictions.
  • Analyze whether model recovers known chemical principles (electronegativity, radius ratios, etc.).

• Chemical Rule Extraction:

  • Cluster materials in the embedding space and analyze synthesizability trends.
  • Identify decision boundaries between synthesizable/unsynthesizable regions.
  • Compare learned "chemical rules" with traditional heuristics.

• Case Study Analysis:

  • Select specific material families for detailed analysis (e.g., perovskites, Heuslers).
  • Trace model predictions to specific training examples using attention mechanisms.
  • Validate whether model captures domain-specific synthesizability factors.

• Visualization:

  • Create low-dimensional projections (t-SNE, UMAP) of material embeddings colored by synthesizability.
  • Generate partial dependence plots for key elemental characteristics.
  • Visualize attention weights in transformer-based models.

Applications and Validation

Performance Benchmarks and Experimental Validation

Machine learning approaches leveraging atomistic representations have demonstrated superior performance across multiple benchmarks:

• Head-to-Head Expert Comparison: In a direct material discovery comparison, SynthNN outperformed all 20 expert materials scientists, achieving 1.5× higher precision while completing the task five orders of magnitude faster than the best human expert [1].

• Temporal Validation: When trained on materials discovered before 2015 and tested on compounds added to databases after 2019, the synthesizability score (SC) model achieved 88.6% true positive rate accuracy, successfully identifying newly synthesizable materials [10].

• Chemical Validity: The BLMM Crystal Transformer generates chemically valid compositions with 89.7% charge neutrality and 84.8% balanced electronegativity—more than four and eight times higher, respectively, compared to pseudo-random sampling [7].

Integration with Materials Discovery Workflows

Advanced synthesizability prediction integrates seamlessly into computational materials screening pipelines:

• Generative Design: Use synthesizability predictions as constraints or objectives in generative models to ensure synthetic accessibility of proposed compositions [7].

• High-Throughput Screening: Apply rapid synthesizability filters to computationally generated hypothetical materials before resource-intensive DFT calculations [10].

• Experimental Prioritization: Rank candidate materials by synthesizability score to focus experimental efforts on the most promising candidates [1].

• Tinkering Design: Utilize blank-filling language models to suggest chemically plausible element substitutions for known materials [7].

The limitations of traditional synthesizability metrics like charge-balancing and formation energy calculations necessitate more sophisticated, data-driven approaches. Atomistic representation learning methods, particularly those based on atom2vec and related techniques, provide a powerful framework for synthesizability prediction that captures complex chemical relationships beyond simple heuristics. The protocols outlined in this Application Note enable researchers to implement these advanced methods, significantly improving the efficiency and success rate of computational materials discovery. As these approaches continue to evolve, integrating synthesizability prediction directly into generative design workflows will further accelerate the identification of novel, synthetically accessible materials for technological applications.

The advent of machine learning (ML) in materials science has shifted the research paradigm from reliance solely on empirical rules and physical simulations to data-driven discovery. Central to this transformation are large-scale, curated databases that provide the foundational data for training and validating predictive models. Within the specific context of developing atomistic representations like atom2vec for predicting chemical synthesizability, three databases play particularly critical roles: the Inorganic Crystal Structure Database (ICSD), the Materials Project (MP), and PubChem. These databases collectively provide comprehensive coverage of known inorganic crystals, computationally characterized materials, and organic molecules, respectively. This application note details the quantitative contributions, experimental protocols, and integrative workflows for leveraging these databases to train and benchmark synthesizability models, providing a practical guide for researchers and scientists in drug development and materials informatics.

Database Profiles and Quantitative Comparison

The table below summarizes the core attributes and primary applications of the three key databases in the context of atom2vec and synthesizability research.

Table 1: Key Databases for Atomistic Model Training

Database Name	Primary Content & Scope	Key Metrics and Volume	Role in atom2vec/Synthesizability Research
Inorganic Crystal Structure Database (ICSD) [1]	Experimentally synthesized and structurally characterized inorganic crystalline materials.	A "nearly complete history" of reported inorganic crystals; used to train models on the entire space of synthesized compositions [1].	Serves as the primary source of positive examples (synthesizable materials) for training supervised and positive-unlabeled (PU) learning models like SynthNN [1].
Materials Project (MP) [12] [3]	A computational database of DFT-calculated material properties, including crystal structures, formation energies, and stability metrics for hundreds of thousands of materials.	Contains data on "hundreds of thousands" of materials; provides standardized formation energies and energies above the convex hull (Ehull) for stability assessment [12] [13].	Provides stability descriptors (e.g., Ehull) used as features or validation metrics. Supplies hypothetical structures for discovery campaigns and seeds prototype-based structure generation [12] [13].
PubChem [14]	An open archive of chemical substances, focusing on small molecules and their biological activities.	Contains over 46 million compound records (as of 2013). A 2015 analysis identified 28,462,319 unique atom environments within its datasets [14].	Provides a vast corpus for unsupervised learning of atom embeddings. The concept of atom environments is directly transferable to learning representations for inorganic solids [15] [14].

Experimental Protocols for Database Utilization

Protocol 1: Training a Synthesizability Classifier (e.g., SynthNN) with ICSD

This protocol outlines the steps for using the ICSD to train a deep learning model for synthesizability prediction, as demonstrated by SynthNN [1].

• Data Acquisition and Curation:

  • Obtain the full ICSD database or a relevant subset.
  • Extract the chemical formulas of all crystalline inorganic materials, representing the set of synthesized (positive) examples.
  • Apply necessary pre-processing, such as removing duplicates and handling non-stoichiometric entries.

• Generation of Unsynthesized Examples:

  • Create a set of artificially generated chemical formulas that are not present in the ICSD. These serve as the unsynthesized (or unlabeled) class in the dataset.
  • The ratio of artificially generated formulas to synthesized formulas (referred to as N_synth) is a critical hyperparameter [1].

• Model Training with Positive-Unlabeled (PU) Learning:

  • Implement a semi-supervised learning approach that treats the unsynthesized materials as unlabeled data.
  • Probabilistically reweight the unlabeled examples according to their likelihood of being synthesizable to account for the fact that some may be synthesizable but not yet discovered [1].
  • Employ an atom2vec-style embedding layer as the input to a neural network (SynthNN). This layer learns optimal vector representations for each atom directly from the distribution of chemical formulas.
  • Train the model to classify formulas as synthesizable or not.

• Model Validation:

  • Benchmark model performance against baseline methods, such as random guessing and the charge-balancing heuristic.
  • Evaluate using standard metrics like precision and recall, while acknowledging that the "unsynthesized" class may contain false negatives [1].

Protocol 2: Integrating DFT Stability from the Materials Project into Predictive Models

This protocol describes how to incorporate thermodynamic stability data from the Materials Project to enhance synthesizability predictions [13].

• Feature Extraction from MP:

  • For a given list of chemical compositions, query the Materials Project API to retrieve computed properties.
  • Extract key stability metrics, most importantly the energy above the convex hull (Ehull), which describes a compound's zero-kelvin thermodynamic stability.
  • Extract or calculate other relevant compositional or structural features available through MP or associated tools like pymatgen and matminer.

• Model Training with Combined Features:

  • Combine the DFT-derived stability features (e.g., Ehull) with composition-based features (e.g., atom2vec embeddings or features from featurizers like MAGPIE or JARVIS).
  • Train a machine learning classifier (e.g., ensemble methods or neural networks) to predict synthesizability, using reported materials (e.g., from ICSD) as the positive class.
  • The model learns the complex relationship where low Ehull is generally indicative of synthesizability, but also accounts for metastable synthesizable materials and stable-yet-unsynthesized materials [13].

• Screening and Discovery:

  • Apply the trained model to screen large sets of hypothetical compounds.
  • Identify promising candidates that are predicted to be synthesizable, which may include both stable and metastable compositions, thereby going beyond a simple Ehull filter [13].

Protocol 3: Learning Atom Embeddings (atom2vec) from a Materials Corpus

This protocol is based on the original atom2vec methodology, which can be applied to a database of material compositions like those in the ICSD or PubChem [15].

• Corpus Construction:

  • Compile a large list of chemical formulas from a chosen database (e.g., all entries in ICSD or a subset of PubChem). This list is the "corpus" analogous to a corpus of text documents.

• Defining Atom Environments:

  • For each chemical formula, generate atom-environment pairs.
  • In a simplified composition-based approach, for a compound like Bi2Se3, the environment for atom Bi is represented as (2)Se3, meaning two central Bi atoms are surrounded by three Se atoms in the remainder of the compound [15].

• Building the Atom-Environment Matrix:

  • Construct a co-occurrence matrix X, where each entry Xij represents the frequency of the i-th atom type appearing in the j-th type of environment across the entire corpus.

• Dimensionality Reduction:

  • Apply a model-free machine learning method, such as Singular Value Decomposition (SVD), to the reweighted and normalized atom-environment matrix.
  • The row vectors corresponding to atoms in the subspace of the d largest singular values become the d-dimensional atom embeddings [15].
  • These embeddings can be clustered and visualized to show that they capture fundamental chemical properties and periodic trends without prior human knowledge [15].

Workflow Visualization: From Databases to Synthesizability Prediction

The following diagram illustrates the integrative workflow for using these databases to train an atom2vec-informed synthesizability model.

Table 2: Essential Computational Tools and Datasets for Synthesizability Research

Tool/Resource Name	Type	Primary Function in Workflow
ICSD [1]	Database	The definitive source for experimentally verified inorganic crystal structures; provides the ground truth for "synthesized" materials.
Materials Project (MP) [12] [3]	Database	Provides pre-computed quantum mechanical properties (formation energy, Ehull) for hundreds of thousands of materials, essential for stability-informed models.
PubChem [14]	Database	A vast source of molecular structures and atom environments, useful for training general-purpose atom embeddings.
atom2vec [15]	Algorithm / Representation	An unsupervised method for learning vector representations of atoms that capture their chemical properties based on co-occurrence in a database.
pymatgen [12]	Python Library	A robust library for materials analysis; used for parsing crystal structures, analyzing phase stability, and integrating with MP data.
matminer [12] [3]	Python Library	An open-source toolkit for data mining in materials science; used to generate a wide array of composition-based and structure-based numerical features.
Positive-Unlabeled (PU) Learning [1]	Machine Learning Framework	A semi-supervised classification approach critical for handling the lack of confirmed negative examples (unsynthesizable materials) in synthesizability prediction.

The discovery and development of new materials are fundamental to technological progress in fields ranging from renewable energy to drug development. A pivotal challenge in this endeavor is accurately predicting material synthesizability—determining whether a hypothetical chemical compound can be successfully realized in the laboratory. Traditional methods, which often rely on proxy metrics like charge-balancing or density functional theory (DFT) calculations, face significant limitations; for instance, charge-balancing criteria alone fail to identify over 60% of known synthesized inorganic materials [1].

Inspired by breakthroughs in natural language processing (NLP), a new paradigm has emerged: representing chemical elements as distributed vector embeddings learned from large-scale data. This approach allows machine learning models to capture complex, multifaceted chemical relationships that are difficult to codify through manual feature engineering. Techniques such as atom2vec and SkipAtom demonstrate that the statistical patterns of "co-occurrence" in materials databases or scientific text can yield powerful, meaningful representations of atoms, mirroring how NLP models like Word2Vec learn semantic meaning from word co-occurrence in text corpora [4] [16]. These learned representations form the foundation for highly accurate predictive models of synthesizability and material properties, enabling more efficient and reliable computational screening of novel materials [1].

Theoretical Foundations: From Word Embeddings to Atom Embeddings

Core NLP Concepts and Their Chemical Analogies

The application of NLP principles to materials science relies on a direct conceptual mapping between linguistic and chemical domains.

• Words and Atoms: In NLP, words are the basic units of meaning. In materials informatics, atoms serve an analogous role as the fundamental building blocks [4].
• Documents and Materials: A document is a structured sequence of words. Similarly, a crystalline material, defined by its chemical formula and atomic structure, can be viewed as a "document" where atoms appear in specific structural contexts [4].
• Vocabulary and The Periodic Table: The set of all unique words in a corpus is the vocabulary. In chemistry, the periodic table of elements constitutes the fundamental vocabulary [16].
• Semantic Similarity and Chemical Similarity: In language, words with similar meanings appear in similar contexts (the "distributional hypothesis"). In chemistry, a parallel principle holds: atoms with similar chemical properties will be found in similar structural environments within materials [4]. For example, calcium and strontium, being chemically similar, might both coordinate with oxygen in comparable geometric patterns.

The Skip-gram Model and its Chemical Adaptations

The Skip-gram model, a cornerstone of modern NLP, aims to predict the context words surrounding a given target word within a predefined window. This training objective forces the model to learn vector embeddings that place words with similar contexts close together in a high-dimensional space [4].

This model has been directly adapted for materials science in two primary ways:

• SkipAtom: This method replaces the textual corpus with a database of crystal structures. Each material is represented as a graph where atoms are nodes connected by edges based on their proximity (e.g., derived from Voronoi decomposition). The training objective is modified to predict the neighboring atoms of a target atom within this crystal graph. The model learns to maximize the average log probability defined as: $$\frac{1}{| M| }\mathop{\sum}\limits_{m\in M}\mathop{\sum}\limits_{a\in {A}_{m}}\mathop{\sum}\limits_{n\in N(a)}\log p(n| a)$$ where M is the set of materials, A_m is the set of atoms in material m, and N(a) are the neighbors of atom a [4].

• atom2vec: This approach uses a similar graph-based context prediction but was notably used to autonomously reproduce the structure of the periodic table. By analyzing the known chemical compounds of 118 elements, the algorithm learned vector embeddings that grouped elements by their chemical properties without any prior chemical knowledge [16].

Advanced Embeddings: From Global to Local Contexts

Recent advancements have extended these core ideas to create more nuanced representations. The Element2Vec framework, for example, processes textual descriptions of elements from sources like Wikipedia using large language models (LLMs). It generates two types of embeddings:

• Global Embedding: A single, general-purpose vector representation for an element derived from its entire text corpus [6].
• Local Embeddings: A set of attribute-specific vectors, each tailored to highlight a particular characteristic (e.g., atomic, chemical, or physical properties) [6].

This approach moves beyond co-occurrence in crystal structures to incorporate rich, descriptive knowledge from scientific literature, providing a more holistic representation of chemical elements.

Experimental Protocols and Application Notes

This section provides detailed methodologies for implementing and applying atom representation learning models in synthesizability research.

Protocol 1: Training a SkipAtom Model

Objective: To learn distributed representations of atoms from a database of crystalline structures.

Materials and Input Data:

• Database: A curated collection of crystal structures, such as the Inorganic Crystal Structure Database (ICSD) or the Materials Project [4].
• Software: A computing environment with Python and machine learning libraries (e.g., PyTorch or TensorFlow).

Methodology:

• Graph Construction: For each crystal structure in the database, generate an undirected graph.
  • Nodes: Represent individual atoms.
  • Edges: Connect atoms that are considered neighbors. A robust method for this is Voronoi decomposition, which identifies nearest neighbors using solid angle weights to determine coordination environments [4].
• Training Pair Generation: Traverse each graph. For every atom (the target), compile all its directly connected neighboring atoms (the context) into (target, context) pairs.
• Model Setup: Implement a shallow neural network with:
  • An input layer that accepts a one-hot encoded vector of the target atom.
  • A single hidden layer (the projection layer) with a linear activation.
  • An output layer with a softmax activation that predicts the probability distribution over all possible context atoms.
• Model Training: Train the network using Maximum Likelihood Estimation to minimize the cross-entropy loss between the predicted context probabilities and the true context atoms. The rows of the projection layer's weight matrix become the final atom embeddings [4].

Protocol 2: Building a Synthesizability Predictor (SynthNN)

Objective: To predict the synthesizability of an inorganic crystalline material given only its chemical formula.

Materials and Input Data:

• Positive Data: Chemical formulas of known synthesized materials, sourced from the ICSD [1].
• Negative Data: Artificially generated chemical formulas that are presumed to be unsynthesized. The ratio of artificial to synthesized formulas (N_synth) is a key hyperparameter [1].
• Model: A deep learning model that uses learned atom embeddings.

Methodology:

• Data Preparation and Representation:
  • Represent each chemical formula using an atom embedding matrix, where each element in the formula is represented by its learned vector (e.g., from SkipAtom or atom2vec). This matrix is optimized alongside other model parameters [1].
  • This approach allows the model to learn an optimal representation of chemical formulas directly from the distribution of synthesized materials, without relying on handcrafted features [1].
• Model Training with Positive-Unlabeled Learning:
  • This is a Positive-Unlabeled (PU) learning problem. Artificially generated materials are treated as unlabeled data rather than definitive negatives, as some may be synthesizable but not yet discovered.
  • The model, often called SynthNN, is trained to classify materials as synthesizable. Unlabeled examples are probabilistically reweighted according to their likelihood of being synthesizable [1].
• Validation: Benchmark the model's precision and recall against baseline methods like random guessing and charge-balancing. SynthNN has been shown to identify synthesizable materials with significantly higher precision than formation energy calculations from DFT [1].

Workflow Visualization

The following diagram illustrates the integrated workflow from data processing to synthesizability prediction, combining the concepts from the protocols above.

Diagram 1: Integrated workflow for atom representation learning and synthesizability prediction.

Benchmarking and Data Presentation

The performance of models leveraging atom embeddings is benchmarked against traditional methods and other feature sets. The following table summarizes key quantitative results from the literature.

Table 1: Benchmarking Performance of Different Models and Featurizers on Material Informatics Tasks

Model/Featurizer	Core Principle	Key Performance Metric	Result	Reference / Context
SynthNN (with atom embeddings)	Learns synthesizability from data of all synthesized materials using atom embeddings.	Precision in identifying synthesizable materials	7x higher precision than DFT-calculated formation energies; outperformed 20 human experts.	[1]
Charge-Balancing Baseline	Predicts synthesizability if a material has a net neutral ionic charge.	Coverage of known synthesized inorganic materials	Correctly identifies only ~37% of known synthesized materials.	[1]
MatterGen (Generative Model)	Diffusion model generating stable, diverse inorganic materials.	Percentage of generated structures that are Stable, Unique, and New (SUN)	>75% of generated structures are stable; 61% are new materials.	[17]
Composition & Structure Featurizer (SAF+CAF)	Generates explainable compositional and structural features for ML.	F1-Score for classifying AB intermetallic crystal structures	0.983 (XGBoost), comparable to other advanced featurizers.	[3]
SOAP Featurizer	Smooth Overlap of Atomic Positions; high-dimensional structural descriptor.	F1-Score for classifying AB intermetallic crystal structures	0.983 (XGBoost), but with 6,633 features (computationally expensive).	[3]

The Scientist's Toolkit: Research Reagent Solutions

This section details the essential computational tools and data resources that form the modern materials informatics pipeline.

Table 2: Key Resources for Atom Representation Learning and Synthesizability Prediction

Tool / Resource	Type	Primary Function	Relevance to Synthesizability Research
Inorganic Crystal Structure Database (ICSD)	Database	A comprehensive collection of published crystal structures.	Serves as the primary source of "positive" data (known synthesized materials) for training models like SynthNN [1].
atom2vec / SkipAtom	Algorithm	Learns atomic embeddings from material structures using NLP-inspired models.	Generates the foundational vector representations of atoms that capture chemical similarity, which are input for predictors [1] [4].
SynthNN	Predictive Model	A deep learning classifier for material synthesizability.	The end-stage model that uses atom embeddings to directly predict the likelihood of a material being synthesizable [1].
mat2vec	Algorithm / Embeddings	Learns atom and material embeddings from scientific literature abstracts.	Provides an alternative, text-based representation of atoms, enriching feature sets [3].
Composition Analyzer Featurizer (CAF)	Software Tool	Generates numerical compositional features from a chemical formula.	Creates human-interpretable features for building explainable ML models, complementing learned embeddings [3].
Positive-Unlabeled (PU) Learning	Computational Framework	A semi-supervised learning paradigm for datasets with only positive and unlabeled examples.	Critical for handling the lack of definitive negative examples (proven unsynthesizable materials) in synthesizability prediction [1].
Fgfr-IN-5	Fgfr-IN-5, MF:C25H22N6O3, MW:454.5 g/mol	Chemical Reagent	Bench Chemicals
Antitumor agent-184	Antitumor agent-184, MF:C22H16N4O2S, MW:400.5 g/mol	Chemical Reagent	Bench Chemicals

The integration of NLP principles into materials science has catalyzed a fundamental shift in how we represent and reason about chemical elements. By treating atoms as words and materials as documents, techniques like atom2vec and SkipAtom automatically learn rich, distributed representations that encapsulate profound chemical relationships. These embeddings have proven to be powerful features for downstream predictive tasks, most notably in addressing the critical challenge of predicting material synthesizability.

Framed within the broader context of atom2vec representation for synthesizability research, this approach demonstrates a significant advantage over traditional methods. Models like SynthNN, built upon these learned embeddings, not only achieve superior precision but also learn foundational chemical principles like charge-balancing and chemical family relationships directly from data [1]. As these representation learning techniques continue to evolve, incorporating multimodal information from text and structure, they pave the way for more reliable, efficient, and accelerated discovery of novel, synthesizable materials.

Building and Deploying Synthesizability Predictors: From SynthNN to DeepSA

The discovery of novel inorganic crystalline materials is a fundamental driver of technological innovation. However, a significant bottleneck exists in transitioning from computationally predicted materials to those that can be experimentally realized. The challenge of predicting synthesizability—determining whether a hypothetical chemical composition can be synthesized as a crystalline solid—remains a critical unsolved problem in materials science [18]. Traditional proxies for synthesizability, such as charge-balancing criteria and thermodynamic stability calculated via density functional theory (DFT), have proven inadequate. Charge-balancing alone identifies only 37% of known synthesized materials, while DFT-based formation energies fail to account for kinetic stabilization and synthetic accessibility [18]. This limitation has created an urgent need for more accurate and efficient predictive methods that can keep pace with high-throughput computational material discovery.

Within this context, the SynthNN model represents a paradigm shift in synthesizability prediction. Developed as a deep learning classification model, SynthNN leverages the entire corpus of known inorganic chemical compositions to directly predict synthesizability from chemical formulas alone, without requiring structural information [18] [19]. By reformulating material discovery as a synthesizability classification task, this approach achieves a critical objective: enabling rapid screening of billions of candidate materials to identify those most likely to be synthetically accessible, thereby increasing the reliability of computational material screening workflows [18].

Architectural Foundation: The atom2vec Representation

At the core of SynthNN's innovative approach is its use of the atom2vec representation framework, which provides a learned, distributed representation of chemical elements [18]. This framework moves beyond traditional fixed chemical descriptors to create an adaptive, data-driven representation optimized specifically for synthesizability prediction.

atom2vec Implementation in SynthNN

The atom2vec framework represents each chemical formula through a learned atom embedding matrix that is optimized alongside all other parameters of the neural network during training [18]. In this architecture:

• Element Embeddings: Each chemical element is represented as a dense vector in a continuous space, with the dimensionality treated as a hyperparameter optimized during model development.
• Composition Representation: Chemical formulas are processed through these embeddings to create a unified representation that captures complex compositional relationships.
• Joint Optimization: The element representations are not pre-trained but are learned jointly with the synthesizability classification task, allowing the model to discover element relationships most relevant to synthesizability.

Remarkably, despite having no explicit chemical knowledge programmed into it, SynthNN learns fundamental chemical principles through this representation. Experimental analyses indicate that the model internalizes concepts of charge-balancing, chemical family relationships, and ionicity directly from the distribution of synthesized materials, utilizing these learned principles to generate synthesizability predictions [18] [20].

Comparative Advantage Over Traditional Representations

The atom2vec framework provides significant advantages over traditional chemical representations:

• Flexibility: Unlike fixed oxidation state tables or manually engineered features, the learned embeddings can adapt to capture complex, non-intuitive patterns in the data.
• Composition Focus: By operating solely on chemical compositions, SynthNN can evaluate materials for which atomic structures are unknown—a critical capability for genuine discovery of novel materials.
• Data-Driven Insights: The model learns which elemental properties and relationships best predict synthesizability without human bias in feature selection.

Model Design and Training Methodology

Neural Network Architecture and Workflow

SynthNN implements a deep learning architecture designed specifically for processing chemical compositions represented through atom2vec embeddings. The model follows a structured workflow from input chemical formula to synthesizability classification, illustrated in the following diagram:

The architectural workflow begins with a chemical formula as input, which is processed through the atom2vec embedding layer to create distributed representations of the constituent elements. These embeddings then pass through multiple feature learning layers that capture complex interactions between elements in the composition. Finally, the classification layer generates a synthesizability probability score, indicating the model's confidence that the input formula can be successfully synthesized [18] [19].

Training Framework: Positive-Unlabeled Learning

A fundamental challenge in synthesizability prediction is the lack of confirmed negative examples—while successfully synthesized materials are documented in databases like the Inorganic Crystal Structure Database (ICSD), failed synthesis attempts are rarely reported [18] [21]. SynthNN addresses this through a positive-unlabeled (PU) learning approach, a semi-supervised learning paradigm that treats unsynthesized materials as unlabeled rather than definitively unsynthesizable [18].

The training process involves:

• Positive Examples: 70,120 synthesized crystalline inorganic materials extracted from the ICSD [18].
• Artificial Negative Examples: Artificially generated unsynthesized materials, treated as unlabeled data in the PU learning framework.
• Probabilistic Reweighting: The model probabilistically reweights unlabeled examples according to their likelihood of being synthesizable, handling the inherent uncertainty in negative example labeling [18].

The ratio of artificially generated formulas to synthesized formulas used in training (referred to as Nâ‚›ynth) is treated as a key hyperparameter, with detailed analysis provided in the supplementary materials of the original publication [18].

Performance Analysis and Benchmarking

Quantitative Performance Metrics

SynthNN demonstrates superior performance compared to traditional synthesizability assessment methods, as quantified through comprehensive benchmarking. The table below summarizes key performance metrics across different prediction approaches:

Table 1: Performance comparison of synthesizability prediction methods

Method	Precision	Recall	Key Advantages	Limitations
SynthNN (threshold=0.5)	0.563	0.604	7× higher precision than DFT; learns chemical principles	Requires training data [18] [19]
Charge-Balancing	0.37 (on known materials)	N/A	Chemically intuitive; computationally simple	Only identifies 37% of known materials [18]
DFT Formation Energy	~0.08 (7× lower than SynthNN)	~0.50	Physics-based; well-established	Misses kinetic stabilization; computationally expensive [18]
Human Experts (best performer)	1.5× lower than SynthNN	N/A	Domain knowledge; contextual understanding	5 orders of magnitude slower than SynthNN [18]

The precision-recall tradeoff for SynthNN can be modulated by adjusting the classification threshold, enabling users to optimize for either high-recall exploration or high-precision targeted discovery:

Table 2: SynthNN performance at different classification thresholds

Threshold	Precision	Recall
0.10	0.239	0.859
0.30	0.419	0.721
0.50	0.563	0.604
0.70	0.702	0.483
0.90	0.851	0.294

Comparative Analysis with Alternative Approaches

Recent advances in synthesizability prediction have introduced several alternative methodologies. The Crystal Synthesis Large Language Models (CSLLM) framework achieves 98.6% accuracy in predicting synthesizability of 3D crystal structures with known atomic positions, significantly outperforming thermodynamic and kinetic stability metrics [22]. However, CSLLM requires complete structural information, limiting its application to materials with known or predicted crystal structures. In contrast, SynthNN's composition-based approach enables screening of entirely novel chemical spaces where structural data is unavailable [18] [22].

Other PU learning approaches for solid-state synthesizability prediction have demonstrated capability in specialized domains. For ternary oxides, a PU learning model trained on human-curated literature data successfully identified 134 likely synthesizable compositions from 4,312 hypothetical candidates [21]. These specialized models benefit from high-quality, domain-specific training data but lack the generalizability of SynthNN across the entire inorganic composition space.

Practical Implementation Protocols

Experimental Setup and Research Reagents

Implementing SynthNN for material discovery workflows requires specific computational resources and data sources, as detailed in the following research reagents table:

Table 3: Essential research reagents for SynthNN implementation

Reagent Solution	Function	Source/Specification
ICSD Data	Source of positive training examples; validation	Inorganic Crystal Structure Database [18]
Pre-trained SynthNN Weights	Model initialization for prediction	Official GitHub Repository [19]
Artificial Negative Generator	Generation of unlabeled examples	Custom implementation per [18]
atom2vec Embeddings	Chemical formula representation	Learned during training [18]
Python/PyTorch Stack	Model training and inference environment	Standard deep learning framework [19]

Step-by-Step Prediction Protocol

For researchers applying SynthNN to screen candidate materials, the following protocol provides a standardized approach:

Protocol 1: Synthesizability Screening of Novel Compositions

• Input Preparation

  • Format chemical formulas using standard notation (e.g., "CsCl", "BaTiO3")
  • Ensure formulas represent charge-neutral compositions
  • Compose candidate list from generative models or high-throughput computations

• Model Inference

  • Load pre-trained SynthNN model from official repository
  • Process formulas through atom2vec embedding layer
  • Forward propagate through trained network architecture
  • Extract synthesizability probabilities from output layer

• Result Interpretation

  • Apply threshold appropriate for discovery objective (0.5 for balanced screening)
  • For high-precision synthesis targeting, use threshold ≥0.7
  • For exploratory discovery of novel systems, use threshold ≤0.3

• Validation Planning

  • Prioritize high-probability candidates for experimental validation
  • Consider chemical feasibility and precursor availability
  • Design synthesis protocols based on analogous synthesized materials

Model Retraining Protocol

For domain-specific applications, researchers may need to retrain SynthNN on specialized materials classes:

Protocol 2: Domain-Specific Model Retraining

• Data Curation

  • Extract domain-specific positive examples from ICSD
  • Generate artificial negative examples matching domain characteristics
  • Maintain appropriate positive-to-unlabeled ratio (default 1:20)

• Model Configuration

  • Initialize with published architecture specifications
  • Set atom2vec embedding dimension (optimized as hyperparameter)
  • Configure learning rate and batch size for stable training

• Training Execution

  • Train for specified epochs with early stopping
  • Validate on held-out set of known materials
  • Tune classification threshold based on precision-recall requirements

• Model Evaluation

  • Benchmark against charge-balancing and DFT formation energy
  • Compare with human expert assessments where feasible
  • Verify learned representations capture domain-specific chemistry

Integration in Materials Discovery Workflows

SynthNN enables a transformative approach to computational materials discovery by integrating synthesizability constraints directly into screening pipelines. The following diagram illustrates this integrated workflow:

This integrated approach addresses the critical bottleneck in materials discovery by prioritizing candidates that balance desirable functional properties with synthetic accessibility. The workflow begins with candidate generation through high-throughput computation or generative models like MatterGen [23], followed by SynthNN screening to filter for synthesizable compositions. The resulting candidates are then prioritized based on both synthesizability scores and predicted functional properties, enabling targeted experimental validation of the most promising materials [18] [23].

In comparative evaluations, this SynthNN-guided approach has demonstrated remarkable effectiveness. In head-to-head material discovery comparisons against 20 expert materials scientists, SynthNN achieved 1.5× higher precision than the best human expert while completing the task five orders of magnitude faster [18]. This demonstrates the transformative potential of integrating data-driven synthesizability prediction into discovery pipelines.

Future Perspectives and Development

The development of SynthNN represents a significant advancement in synthesizability prediction, but several frontiers remain for further development. Future iterations could benefit from multi-modal learning that incorporates both compositional and structural information where available, potentially bridging the gap between composition-based models like SynthNN and structure-based approaches like CSLLM [22]. Additionally, transfer learning approaches could enable effective fine-tuning of the general SynthNN model to specialized material classes with limited domain-specific data, similar to techniques demonstrated for perovskite synthesizability prediction [21].

Another promising direction involves integration with synthesis planning models that suggest specific precursors and reaction conditions. While SynthNN focuses on compositional synthesizability, complementary approaches are emerging for predicting synthetic methods and suitable precursors, with recent models achieving >90% accuracy in classifying synthetic methods and >80% success in identifying appropriate precursors [22]. Combining these capabilities could create a comprehensive pipeline from compositional design to synthesis recipe recommendation.

As autonomous materials discovery platforms continue to develop, SynthNN-class models will play an increasingly critical role in ensuring that computationally discovered materials are synthetically accessible, ultimately accelerating the translation of predicted materials into realized technological innovations.

Positive-Unlabeled (PU) learning is a growing subfield of machine learning that addresses the challenge of binary classification when explicit negative training data is unavailable. In contrast to conventional supervised models trained on both positive and negative data, a PU learning algorithm works on a training set containing only labeled positive instances and unlabeled instances that could be positive or negative [24]. This approach is particularly effective in real-world scenarios where negative examples are missing, difficult to define, or extremely diverse. The core problem PU learning solves is training a accurate binary classifier without access to explicitly labeled negative data, making it a powerful tool for many practical applications where negative information is inherently absent [24].

PU learning is fundamentally important in domains like medical diagnosis, where a patient's medical records list only diagnosed diseases (positives) but cannot comprehensively list all conditions the patient does not have. Similarly, in fake comment detection, systems often identify only definite fake comments (positives) without a clean set of confirmed real comments [24]. A key development in modern PU learning is the shift from instance-independent assumptions, where positive data is randomly labeled, to instance-dependent scenarios, where the likelihood of a positive instance being labeled depends on its specific features, such as its distance from a potential decision boundary [24].

PU Learning in Chemical and Materials Science

The application of PU learning has proven particularly valuable in chemical and materials science, where it helps navigate the vast, unexplored regions of chemical space. A central challenge in materials discovery is predicting the synthesizability of a material—whether it can be synthetically accessed through current methods, regardless of whether it has been reported before [1]. Traditional proxies for synthesizability, such as charge-balancing or thermodynamic stability calculated from Density Functional Theory (DFT), often show limited accuracy. For instance, charge-balancing alone identifies only 37% of known synthesized inorganic materials, a figure that drops to just 23% for known ionic binary cesium compounds [1].

In this context, PU learning provides a robust framework for predicting synthesizability by treating all known synthesized materials as the positive set and a vast space of hypothetical, computer-generated compositions as the unlabeled set. This unlabeled set contains mostly unsynthesizable materials but may also include synthesizable ones that have not yet been discovered or made [1]. This approach was successfully implemented in the SynthNN (Synthesizability Neural Network) model, which directly predicts the synthesizability of inorganic chemical formulas without requiring structural information [1].

The Role of atom2vec Representation

A critical innovation in applying PU learning to materials science is the use of the atom2vec representation. The SynthNN model leverages this representation to learn an optimal, data-driven featurization of chemical formulas [1]. Instead of relying on human-engineered features or assumed synthesizability principles, atom2vec represents each chemical formula by a learned atom embedding matrix that is optimized alongside all other parameters of the neural network [1].

In this framework, the chemistry of synthesizability is learned directly from the distribution of all previously synthesized materials. The dimensionality of this representation is a hyperparameter, and the model does not require pre-defined assumptions about factors influencing synthesizability [1]. This allows the model to automatically capture complex chemical principles such as charge-balancing, chemical family relationships, and ionicity, utilizing them to generate accurate synthesizability predictions [1]. This capability demonstrates how atom2vec provides a powerful and adaptive foundation for representing chemical knowledge in PU learning tasks.

Quantitative Performance of PU Learning Methods

The performance of PU learning models, particularly in synthesizability prediction, can be quantitatively compared against traditional baselines. The following table summarizes key performance metrics for different approaches as demonstrated on benchmark tasks.

Table 1: Performance Comparison of Synthesizability Prediction Methods

Method	Key Principle	Reported Performance Advantage	Applicability
SynthNN (with atom2vec) [1]	Deep learning model using atom2vec representation on all synthesized materials data	7x higher precision than DFT-based formation energy; 1.5x higher precision than best human expert	Direct synthesizability prediction from composition
Charge-Balancing Baseline [1]	Filters materials based on net neutral ionic charge using common oxidation states	Identifies only 37% of known synthesized materials	Simple heuristic filter
PSPU Framework [25]	Uses PU model to generate pseudo-supervision and applies non-PU objectives for correction	Outperforms recent PU methods on MNIST, CIFAR-10, CIFAR-100 in balanced/imbalanced settings	General vision tasks and industrial anomaly detection
NAPU-bagging SVM [26]	Ensemble SVM on bags of positive, negative, and unlabeled data to manage false positive rates	Manages false positive rates while maintaining high recall rates; identifies novel multi-target drug hits	Virtual screening for multi-target drug discovery

Beyond synthesizability, PU learning methods have shown strong results in other domains. The recently proposed PSPU framework significantly outperforms recent PU learning methods on standard datasets like MNIST, CIFAR-10, and CIFAR-100 under both balanced and imbalanced settings [25]. Furthermore, the Negative-Augmented PU-bagging (NAPU-bagging) SVM has demonstrated capability in managing false positive rates while maintaining high recall, which is crucial for virtual screening in multi-target drug discovery [26].

Detailed Experimental Protocols

Protocol 1: Implementing an atom2vec-based Synthesizability Predictor

This protocol details the procedure for developing a synthesizability prediction model using atom2vec representation and PU learning, based on the SynthNN methodology [1].

• Step 1: Data Curation and Preparation

  • Positive Set Construction: Extract known synthesized inorganic crystalline materials from the Inorganic Crystal Structure Database (ICSD). This set constitutes the definitive positive examples.
  • Unlabeled Set Generation: Generate a large set of hypothetical chemical formulas through combinatorial element substitution or other generative algorithms. This set represents the unlabeled data, containing both unsynthesizable and potentially synthesizable but undiscovered materials.

• Step 2: Data Representation with atom2vec

  • Represent each chemical formula in the dataset as a sequence of elemental symbols, typically sorted by electronegativity.
  • Initialize an embedding matrix for all elements in the periodic table. The dimensionality of this matrix is a key hyperparameter.
  • Process the sequences through the model, allowing the atom2vec embeddings to be optimized jointly with the rest of the neural network parameters during training, thereby learning a domain-specific representation.

• Step 3: Model Architecture and PU Training

  • Employ a deep learning architecture, such as a multi-layer perceptron or transformer, that takes the learned atom2vec representations as input.
  • Implement a PU learning objective function. A common choice is a risk estimator that combines the loss on the labeled positive data with an estimated loss on the negative examples extracted from the unlabeled set, often using the class prior probability.
  • Train the model to classify compositions as synthesizable (positive) or not, using the positive set and the unlabeled set.

• Step 4: Model Validation and Testing

  • Evaluate the model on a held-out test set of known synthesized materials (positives) and artificially generated unsynthesized materials (treated as negatives for benchmarking).
  • Acknowledge that the "negative" test set may contain some synthesizable materials, which will lead to a conservative estimate of model precision.

Protocol 2: General Workflow for PU Learning

This protocol outlines a standard, high-level workflow for applying PU learning, adaptable to various domains including materials science and drug discovery [24] [25] [26].

• Step 1: Training Set Generation

  • Choose a data generation scenario based on data structure. The case-control scenario independently samples labeled positives from P(x|y=+1) and unlabeled data from the marginal distribution P(x). The censoring (one-sample) scenario assumes all data is from P(x), with positives being labeled with probability α.

• Step 2: Strategy Selection for Exploiting Unlabeled Data

  • Two-Step Strategy: Identify reliable negative instances from the unlabeled set and then apply supervised learning. This method is highly dependent on accurate identification of negative samples [24].
  • Biased Learning (Cost-sensitive): Treat all unlabeled instances as noisy negatives and reweight training instances to correct for the biased data distribution. This often requires an estimate of the class prior [24].
  • Risk Estimation: Use an unbiased risk estimator that reformulates the standard classification risk using only positive and unlabeled data, avoiding the need for explicit negative examples [25].

• Step 3: Classifier Training

  • Apply the selected strategy to train a binary classifier (e.g., Logistic Regression, Support Vector Machine, or Neural Network) using the positive and unlabeled data.

• Step 4: Iterative Refinement (Advanced)

  • For frameworks like PSPU, an iterative process is used: first, train an initial PU model; second, use it to gather confident pseudo-labels from the unlabeled data; and third, use these pseudo-labels with supervised objectives to refine the model, often with consistency regularization to mitigate noise [25].

Visualizing Workflows and Relationships

The SynthNN Model Workflow with atom2vec

The following diagram illustrates the workflow of the SynthNN model for predicting synthesizability, integrating atom2vec representation and PU learning [1].

SynthNN Synthesizability Prediction Workflow

General PU Learning Process

This diagram outlines the standard process for a Positive and Unlabeled (PU) learning task, from data preparation to model deployment [24] [25].

General PU Learning Process

Table 2: Key Resources for PU Learning in Materials and Drug Discovery

Resource / Tool	Function / Purpose	Example Use Case
Inorganic Crystal Structure Database (ICSD) [1]	Provides a comprehensive collection of known, synthesized inorganic crystal structures used as the positive set.	Curating positive training data for synthesizability prediction models like SynthNN.
atom2vec Representation [1]	Learns an optimal, data-driven numerical representation of chemical formulas from the distribution of known materials.	Featurizing chemical compositions for input into deep learning models in SynthNN.
AiZynthFinder [27]	An open-source tool for computer-aided synthesis planning (CASP); used to validate synthesizability.	Generating ground-truth data for training CASP-based synthesizability scores or validating model outputs.
Urban Institute R Theme (urbnthemes) [28]	An R package that provides pre-defined themes to ensure consistent, accessible, and publication-ready data visualizations.	Creating standardized charts and graphs for reporting model performance and data analysis.
NAPU-bagging SVM [26]	A specific PU-learning algorithm using ensemble Support Vector Machines on resampled data bags.	Virtual screening for multi-target drug discovery, managing false positive rates while maintaining high recall.
Zinc Database [27]	A large database of commercially available chemical compounds, often used as a source of potential building blocks.	Defining the space of accessible starting materials for computer-aided synthesis planning (CASP).

The discovery of new inorganic crystalline materials is fundamental to technological advancement, yet a significant bottleneck exists: reliably predicting whether a hypothetical material is synthesizable. The chemical formula alone is an insufficient predictor, as synthesizability is influenced by a complex interplay of thermodynamic stability, kinetic accessibility, and synthetic pathway feasibility. For decades, charge-balancing, derived from ionic oxidation states, served as a primary, rule-based proxy for synthesizability. However, this approach is remarkably inflexible, failing to account for metallic or covalent bonding and performing poorly quantitatively; it identifies only 37% of known synthesized inorganic materials and a mere 23% of known binary cesium compounds as charge-balanced [1].

The adoption of Density-Functional Theory (DFT) to calculate thermodynamic formation energies represented a major step forward. This method flags materials that are thermodynamically unstable against decomposition to other phases. Nonetheless, its predictive power is limited because it fails to capture kinetic stabilization and non-equilibrium synthetic pathways, resulting in an inability to distinguish synthesizable materials from unsynthesized ones and capturing only about 50% of synthesized inorganic crystalline materials [1].

The field is now undergoing a paradigm shift, moving beyond composition-based descriptors to models that integrate structural motifs. These motifs—recurring, local atomic arrangements like coordination polyhedra or specific chain and ring structures—encode critical chemical intelligence about a material's stability and likely synthesis. This application note details protocols for integrating these structural motifs with advanced graph networks and the AMDNet framework, contextualized within the broader thesis of enhancing atom2vec-based synthesizability predictions.

Theoretical Foundation: From Composition to Structure

Theatom2vecRepresentation and Its Limitations

The atom2vec framework revolutionizes composition-based material discovery by learning a continuous vector representation for each element directly from the distribution of known chemical formulas in massive databases like the Inorganic Crystal Structure Database (ICSD) [1]. This data-driven approach allows a model to infer chemical relationships without pre-defined chemical knowledge. A deep learning synthesizability model (SynthNN) leveraging atom2vec has demonstrated a 7x higher precision in identifying synthesizable materials compared to using DFT-calculated formation energies alone. In a head-to-head discovery challenge, SynthNN outperformed 20 expert material scientists, achieving 1.5x higher precision and completing the task five orders of magnitude faster than the best human expert [1].

Despite its power, a fundamental limitation of atom2vec and other composition-only models is the lack of explicit structural information. A chemical formula, such as SiOâ‚‚, does not distinguish between the vastly different properties and synthesizability of quartz, cristobalite, or amorphous silica glass.

The Critical Role of Structural Motifs

Structural motifs provide the missing link between a chemical formula and a material's realizable crystal structure. They represent the local atomic environment and their connectivity, forming the "building blocks" of crystals. A structure motif–centric, graph-based deep learning framework has been proposed for inorganic crystalline systems, treating these motifs as the fundamental units for analysis and prediction [29]. The central thesis is that integrating these motifs with compositional embeddings will yield a more complete and accurate predictor of synthesizability.

Table 1: Comparative Analysis of Material Representation Approaches

Representation Type	Key Features	Advantages	Limitations
Compositional (e.g., atom2vec)	Learned elemental embeddings from chemical formulas [1].	Computationally lightweight; no structure required; learns chemical trends.	Cannot differentiate between polymorphs; blind to local structure.
Charge-Balancing	Rule-based; checks net neutral ionic charge [1].	Simple, chemically intuitive.	Inflexible; low accuracy (e.g., 37% on ICSD); fails for non-ionic systems.
Global Crystal Graph	Atoms as nodes, bonds as edges [3].	Captures full crystal structure; powerful for property prediction.	Requires a known crystal structure; not for discovery of new compositions.
Structural Motif Graph	Motifs as nodes, connections between motifs as edges [29].	Encodes higher-level chemical intelligence; more interpretable.	Requires motif identification; increased complexity.

Integrated Framework: AMDNet and Graph Networks

The proposed integrated framework leverages the strengths of both compositional and structural representations.

The following workflow diagram outlines the protocol for integrating structural motifs with compositional models for synthesizability prediction.

System Components

Composition Analyzer Featurizer (CAF)

For the composition branch, tools like the Composition Analyzer Featurizer (CAF) can be used. CAF is an open-source Python program that generates numerical compositional features from a list of chemical formulas. It parses formulas into constituent elements and their stoichiometric ratios, calculating features based on element properties (e.g., atomic radius, electronegativity) weighted by stoichiometry. It can generate 133 compositional features, providing a rich vector beyond basic atom2vec embeddings [3].

Structure Analyzer Featurizer (SAF) and Motif Extraction

For the structural branch, the Structure Analyzer Featurizer (SAF) ingests CIF files to generate supercells and extract 94 numerical structural features [3]. The key step is converting this global structure into a motif-centric graph. This involves:

• Identifying Motifs: Using algorithms to detect local coordination environments (e.g., tetrahedra, octahedra) or employing a "group graph" approach that breaks down the structure into functional-group-like substructures (e.g., carbonate groups, silicate tetrahedra) [30] [29].
• Building the Motif Graph: Representing each identified motif as a node. The edges between nodes represent the chemical bonds or connections between these motifs, forming a higher-level graph of the material's architecture [29].

Feature Fusion and AMDNet

The compositional vector from CAF (or atom2vec) and the structural descriptor vector from the motif graph are concatenated into a unified feature vector. This fused vector is then processed by a graph neural network architecture like AMDNet (Atomistic Molecular Deep Network) or a custom Graph Isomorphism Network (GIN). The GIN is particularly powerful as it is theoretically capable of distinguishing different graph structures as effectively as the Weisfeiler-Lehman test, making it well-suited for discerning subtle differences in material topologies [30]. This network learns the complex, non-linear relationships between the composition, local structure, and the final synthesizability classification.

Experimental Protocols

Protocol 1: Building a Synthesizability Classification Model

Application: Training a model to classify hypothetical materials as "synthesizable" or "unsynthesizable." Reagents & Data Sources:

• Primary Data: The Inorganic Crystal Structure Database (ICSD) [1].
• Positive Examples: All synthesized crystalline inorganic materials from ICSD.
• Negative Examples: Artificially generated chemical formulas that are not present in ICSD, acknowledging that some may be synthesizable but are treated as "unlabeled" in a Positive-Unlabeled (PU) learning framework [1].
• Software: Python with libraries like PyTorch or TensorFlow, Pymatgen for structure analysis, and the CAF/SAF featurizers [3].

Methodology:

• Data Curation: Extract ~50,000 unique chemical formulas from ICSD. Generate a larger set (e.g., 500,000) of hypothetical formulas by combining elements in ratios not found in ICSD.
• Featurization:
  • For all formulas, generate compositional features using atom2vec or CAF.
  • For formulas in ICSD, obtain their crystal structures and generate motif-graph features using SAF and motif-extraction algorithms.
  • For hypothetical formulas without a known structure, use a pre-trained Crystal Graph Convolutional Neural Network (CGCNN) to predict a likely crystal structure [3], then extract its motif-graph features.
• Model Training: Split the data into training (80%), validation (10%), and test (10%) sets. Train an AMDNet model using the fused feature vectors to predict a binary synthesizability label. Use the validation set for hyperparameter tuning.
• Validation: Benchmark the model's precision and recall against baseline methods like charge-balancing and DFT formation energy on the held-out test set. The expected outcome is a significant outperformance of these baselines.

Protocol 2: Interpretable Identification of Stabilizing Motifs

Application: Explaining why a specific material is predicted to be synthesizable by identifying critical structural motifs. Reagents & Data Sources:

• A trained synthesizability model from Protocol 1.
• Target material(s) for analysis.
• Explanation techniques like attention mechanisms or GNNExplainer.

Methodology:

• Prediction: Input the target material's formula and predicted structure into the trained model.
• Attention Analysis: If the model uses an attention mechanism, extract the attention weights from the motif-graph processing layer. These weights indicate the relative importance the model assigns to each motif node in the graph for making its final prediction.
• Substructure Importance Mapping: Map the high-attention motifs back to the original crystal structure. Visually highlight these motifs.
• Validation: Correlate the identified high-importance motifs with known chemical knowledge. For example, the model might correctly identify that the stability of a particular perovskite is heavily influenced by the connectivity of its corner-sharing octahedra, a well-established stabilizing feature.

Table 2: Key Research Reagent Solutions for Computational Material Discovery

Research Reagent / Tool	Type	Primary Function	Application in Protocol
Inorganic Crystal Structure Database (ICSD)	Data	Provides a comprehensive set of known, synthesized crystal structures for training and benchmarking.	Source of positive examples and structural data [1].
Composition Analyzer Featurizer (CAF)	Software	Generates 133 human-interpretable numerical features from a chemical formula.	Compositional featurization of both known and hypothetical materials [3].
Structure Analyzer Featurizer (SAF)	Software	Generates 94 numerical structural features from a .cif file by creating a supercell.	Structural featurization of materials with known structures [3].
Graph Isomorphism Network (GIN)	Algorithm	A powerful graph neural network for learning representations of graph-structured data.	Core architecture for processing the fused compositional and motif-graph data [30].
Attention Mechanism	Algorithm	Allows the model to focus on the most relevant parts of the input (e.g., specific motifs).	Enables interpretability by highlighting stabilizing structural motifs [31].

Visualization and Interpretation

The motif-centric approach provides a more chemically intuitive path to model interpretation. As demonstrated in the MMGX (Multiple Molecular Graph eXplainable discovery) framework, using multiple graph representations (atom-level and substructure-level) provides more comprehensive and consistent interpretations, aligning better with a chemist's understanding of functional groups and key substructures [31].

The following diagram illustrates the process of transforming a crystal structure into a motif graph for model interpretation.

The integration of structural motifs with compositional models like atom2vec represents a necessary evolution in the computational prediction of material synthesizability. By moving beyond composition to embrace the rich information encoded in local atomic environments, frameworks like AMDNet coupled with motif-based graph networks achieve higher precision and provide a pathway to chemically interpretable results. The detailed protocols for featurization, model training, and interpretation outlined herein provide researchers with a practical roadmap to implement this advanced approach, accelerating the reliable discovery of novel, synthesizable materials.

The application of artificial intelligence (AI) in drug discovery has revolutionized the process of identifying and designing new therapeutic compounds. A critical challenge in this field is the synthetic accessibility of AI-generated molecules; a compound may be theoretically optimal for binding to a biological target but practically useless if it cannot be synthesized efficiently in a laboratory. The ability to accurately predict the synthesizability of organic compounds is therefore paramount for reducing the time and cost associated with drug development. This application note explores DeepSA, a deep-learning driven predictor designed to assess the synthetic accessibility of chemical compounds. Framed within the broader research on atom2vec representation for chemical formula synthesizability, this document details the application, protocol, and key resources for using DeepSA in drug discovery pipelines, providing researchers and scientists with a practical tool for prioritizing chemically tractable lead compounds.

DeepSA: A Deep-Learning Driven Predictor

DeepSA is a chemical language model that predicts the synthetic accessibility of organic compounds directly from their Simplified Molecular Input Line Entry System (SMILES) representations, a standard string-based notation for molecular structures [32]. By leveraging various natural language processing (NLP) algorithms, DeepSA learns the complex relationship between a molecule's textual representation and its synthesizability. The model was trained on a large dataset of 3,593,053 molecules, enabling it to distinguish between easy-to-synthesize (ES) and hard-to-synthesize (HS) compounds with high accuracy [32]. Its performance is benchmarked by an Area Under the Receiver Operating Characteristic Curve (AUROC) of 89.6%, indicating a high degree of predictive power [32].

Performance Comparison with Existing Methods

DeepSA's performance has been rigorously tested against other state-of-the-art synthesizability assessment tools. The following table provides a quantitative comparison of DeepSA with other popular methods, summarizing their core approaches and key performance metrics on independent test sets.

Table 1: Quantitative Comparison of Synthetic Accessibility Prediction Tools

Method	Core Approach	Training Data Source	Reported AUROC	Output Range/Type
DeepSA	Chemical Language Model (NLP on SMILES)	3.59 million molecules; Retro* & SYBA datasets [32]	89.6% [32]	ES/HS Classification
GASA	Graph Attention Network	Retrosynthesis analysis software (Retro*) [32]	Outperformed by DeepSA [32]	ES/HS Classification
SYBA	Bernoulli Naïve Bayes on molecular fragments	Purchasable molecules & Nonpher-generated molecules [32]	Lower than DeepSA [32]	ES/HS Classification
SAscore	Historical synthesis knowledge & molecular complexity	Millions of synthesized chemicals [32]	Lower than DeepSA [32]	Score (1-10)
SCScore	Deep Neural Network	12 million reactions from Reaxys [32]	Lower than DeepSA [32]	Score (1-5)

This comparison demonstrates that DeepSA provides a significant advantage in discrimination accuracy, helping users select less expensive molecules for synthesis and thereby reducing the time and cost required for drug discovery and development [32].

The following table details the essential computational tools, datasets, and software required for employing DeepSA and related synthesizability assessment methods in a research environment.

Table 2: Essential Research Reagents & Computational Resources

Item Name	Function/Application	Source/Availability
DeepSA Web Server	Online platform for predicting synthesizability from SMILES strings.	Publicly available at: https://bailab.siais.shanghaitech.edu.cn/services/deepsa/ [32]
DeepSA Code	Open-source code for local implementation and customization.	GitHub: https://github.com/Shihang-Wang-58/DeepSA [32]
ChEMBL Database	A manually curated database of bioactive molecules with drug-like properties, used for training and testing.	https://www.ebi.ac.uk/chembl/ [32]
ZINC15 Database	A commercial database of compounds for virtual screening, used as a source of easy-to-synthesize molecules.	https://zinc15.docking.org/ [32]
USPTO Dataset	A massive dataset of chemical reactions used for training retrosynthesis and forward prediction models.	United States Patent and Trademark Office [33]
Retro* Software	A neural-based retrosynthetic planning algorithm used to generate training data by determining synthesis steps.	Software algorithm [32]

Link to Atom2Vec and Material Representation

The broader context of this research involves the use of advanced machine learning to learn meaningful representations of atoms and molecules directly from data. The atom2vec framework is a foundational concept in this area. Inspired by natural language processing techniques, atom2vec learns the properties of atoms as high-dimensional vectors by analyzing the contexts in which they appear across a vast database of known compounds and materials [5] [15]. This unsupervised approach allows the machine to discover periodic trends and chemical similarities without relying on pre-defined human knowledge, such as atomic number or electronegativity [15]. The resulting atom vectors serve as powerful, machine-learned features that can be used as input for predictive models of material properties, including formation energy [15].

While DeepSA operates directly on SMILES strings rather than atom vectors, it is philosophically aligned with the atom2vec paradigm. Both methods demonstrate that machines can learn complex chemical concepts directly from raw data—be it the co-occurrence of atoms in crystal structures (for atom2vec) or the sequence of characters in a SMILES string (for DeepSA). This represents a shift away from feature engineering based on human intuition and towards allowing AI to develop its own informative representations for materials discovery and drug design [15].

Experimental Protocol for Synthesizability Assessment

This section provides a detailed, step-by-step protocol for using the DeepSA model to evaluate the synthetic accessibility of a set of candidate compounds, as might be generated by a molecular generative AI.

Protocol: Using DeepSA for Compound Prioritization

Objective: To classify a library of generated organic compounds as Easy-to-Synthesize (ES) or Hard-to-Synthesize (HS) using the DeepSA predictor.

Principle: The DeepSA model processes the SMILES string of a molecule through its deep neural network architecture, which has been trained to associate specific structural patterns and sequences with synthesizability. The model outputs a classification and a probability score, allowing researchers to filter and prioritize compounds for synthesis.

Materials and Equipment:

• Computer with internet access (for web server) or Python environment (for local code).
• List of candidate compounds represented as valid SMILES strings.
• (Optional) DeepSA software installed locally from GitHub.

Procedure:

• Input Preparation:
  • Compile the list of candidate molecules into a text file, with one SMILES string per line.
  • Ensure all SMILES strings are valid and canonicalized to avoid representation errors.

• Model Interaction:

  • Via Web Server: a. Navigate to the DeepSA web server: https://bailab.siais.shanghaitech.edu.cn/services/deepsa/. b. Upload the prepared text file or copy-paste the SMILES strings into the input field. c. Initiate the prediction job.
  • Via Local Code: a. Load the DeepSA model in your Python script. b. Use the provided functions to read the SMILES file and run the prediction.

• Output and Analysis:

  • The model will return a results file. For each input SMILES, the output will typically include:
    • Predicted Class: Either "ES" (Easy-to-Synthesize) or "HS" (Hard-to-Synthesize).
    • Prediction Score/Probability: A numerical value indicating the confidence of the prediction, often corresponding to the probability of being HS.
  • Prioritize compounds classified as "ES" for further experimental consideration.

• Validation (Recommended):

  • For critical compounds, especially those with borderline prediction scores, consult with a medicinal or synthetic chemist.
  • Consider running the candidate molecules through additional, complementary predictors (e.g., SAscore, SCScore) as a cross-validation step [32].

Troubleshooting Notes:

• Invalid SMILES: Re-generate or correct the SMILES strings using a chemical toolkit (e.g., RDKit).
• Ambiguous Results: If a compound is predicted as HS but is structurally similar to a known drug, investigate the specific structural features flagged as complex (e.g., unusual ring systems, stereochemistry).

Workflow and Dataflow Diagram

The following diagram visualizes the integrated workflow of molecule generation, synthesizability assessment with DeepSA, and the connection to the broader atom2vec representation concept. This provides a logical map of how these components interact in a modern drug discovery pipeline.

Diagram Title: DeepSA in the Drug Discovery Workflow

DeepSA represents a significant advancement in the computational prediction of organic compound synthesis. By providing a highly accurate, deep-learning-based tool, it addresses a critical bottleneck in AI-driven drug discovery. Its integration into the molecular design cycle allows researchers to focus resources on compounds that are not only therapeutically promising but also synthetically feasible. When viewed as part of the larger trend of machine-learned material representations—epitomized by atom2vec—DeepSA underscores a paradigm shift towards data-centric, AI-powered discovery in chemistry and materials science. The provided protocols and resources equip scientists with the necessary knowledge to implement this powerful predictor, thereby accelerating the journey from conceptual design to synthesized candidate.

The traditional drug discovery pipeline, exemplified by the Design-Make-Test-Analyze (DMTA) cycle, is undergoing a significant transformation through the incorporation of artificial intelligence and high-throughput computational methods [27]. A critical challenge limiting the broader adoption of de novo drug design techniques in the "Design" phase is the generation of unrealistic, non-synthesizable molecular structures [27]. While virtual screening platforms have advanced to efficiently screen billion-compound libraries in days [34], this computational efficiency becomes irrelevant if identified hits cannot be practically synthesized in the laboratory.

This application note addresses the integration of synthesizability assessment into high-throughput virtual screening workflows. We focus particularly on the concept of "in-house synthesizability" – predicting whether molecules can be synthesized using a specific, limited collection of available building blocks, rather than assuming infinite commercial availability [27]. This approach bridges the gap between computational screening and practical laboratory constraints, enabling research groups to prioritize candidates that are both biologically active and synthesizable within their resource limitations.

Synthesizability Assessment Methodologies

Computer-Aided Synthesis Planning (CASP)

Computer-Aided Synthesis Planning determines synthesis routes by deconstructing molecules recursively into molecular precursors until a collection of commercially available "building blocks" is identified [27]. Contemporary approaches employ neural networks to encapsulate backward reaction logic and search algorithms to find possible multi-step reaction pathways [27]. While highly accurate, full CASP is computationally intensive, requiring minutes to hours per molecule, making it incompatible with most optimization-based de novo design methods that require numerous optimization iterations [27].

CASP-Based Synthesizability Scores

As a more efficient alternative, CASP-based synthesizability scores approximate synthesis planning results by learning the relationship between a molecule's structure and the successful identification of a synthesis route [27]. These scores are trained on the outcomes of synthesis planning runs and can be formulated as either classification tasks (predicting synthesis planning success) or regression tasks (predicting synthesis route properties) [27]. These learned scores provide a fast measure of synthesizability, making them suitable for post-generation virtual screening or de novo drug design where full CASP would be computationally prohibitive [27].

Positive-Unlabeled Learning for Synthesizability Prediction

For solid-state materials synthesis, positive-unlabeled (PU) learning approaches have shown promise in predicting synthesizability from limited data [21]. This semi-supervised method is particularly valuable given the scarcity of reported failed synthesis attempts in scientific literature. PU learning models trained on human-curated synthesis data from literature can effectively identify synthesizable compositions from hypothetical candidates [21].

Quantitative Analysis of Synthesizability Factors

Impact of Building Block Availability

The transfer of synthesis planning from extensive commercial building block collections to a limited in-house setting was quantitatively evaluated using the AiZynthFinder toolkit [27]. The results demonstrate the feasibility of operating with constrained building block libraries.

Table 1: Synthesis Planning Performance with Different Building Block Libraries

Building Block Source	Number of Building Blocks	Solvability Rate (Caspyrus)	Solvability Rate (ChEMBL)	Average Route Length
Zinc (Commercial)	17.4 million	~70%	~70%	Shorter by ~2 steps
Led3 (In-House)	5,955	~60%	~60%	Longer by ~2 steps

Despite a 3000-fold reduction in available building blocks, the solvability rate decreased by only approximately 12%, though synthesis routes were typically two reaction steps longer on average [27]. This confirms that maintaining an extensive commercial inventory is unnecessary for identifying potential synthesis routes.

Performance of Virtual Screening with Synthesizability Integration

The RosettaVS virtual screening method, which incorporates physics-based binding affinity prediction and receptor flexibility modeling, demonstrates state-of-the-art performance in identifying true binders [34].

Table 2: Virtual Screening Performance Benchmarks

Screening Method	Top 1% Enrichment Factor (CASF2016)	Screening Power (DUD Dataset)	Key Advantage
RosettaGenFF-VS	16.72	Leading performance	Models receptor flexibility
Second-best method	11.90	Lower than RosettaGenFF-VS	Varies by method
Traditional methods	<11.90	Moderate performance	Established use

The RosettaGenFF-VS method achieves a top 1% enrichment factor of 16.72, significantly outperforming the second-best method (EF1% = 11.9) on the CASF2016 benchmark [34]. This demonstrates the critical importance of accurate binding affinity prediction in virtual screening workflows.

Integrated Workflow Protocol

Comprehensive Virtual Screening with Synthesizability Assessment

The following workflow integrates high-throughput virtual screening with practical synthesizability assessment, creating an efficient pipeline from computational screening to laboratory synthesis.

Workflow Diagram 1: Integrated Virtual Screening and Synthesizability Assessment. This workflow demonstrates the seamless integration of computational screening with practical synthesizability evaluation, highlighting the AI-accelerated component that enables efficient processing of ultra-large compound libraries.

Protocol Steps

• Initial Virtual Screening: Screen multi-billion compound libraries using the RosettaVS platform with express (VSX) mode for rapid initial screening. The OpenVS platform employs active learning techniques to train a target-specific neural network during docking computations, efficiently triaging and selecting promising compounds for more expensive docking calculations [34]. Screening can be completed in less than seven days using a high-performance computing cluster with 3000 CPUs and one GPU per target [34].

• Candidate Selection: Select top candidates based primarily on predicted binding affinity and complementary structural diversity to ensure exploration of various chemotypes. At this stage, apply simple synthesizability heuristics (e.g., structural complexity filters) as a preliminary filter [27].

• In-House Synthesizability Scoring: Apply a rapidly retrainable in-house synthesizability score to the candidate compounds. This score predicts whether molecules are synthesizable using the available in-house building block collection (typically 5,000-10,000 compounds) without relying on external commercial resources [27]. Training this score requires a well-chosen dataset of approximately 10,000 molecules, allowing rapid retraining to accommodate changes in building block inventory [27].

• Computer-Aided Synthesis Planning: Perform detailed CASP analysis on synthesizable candidates using tools like AiZynthFinder configured with the specific in-house building block collection [27]. This step verifies synthesizability predictions and generates detailed synthesis routes.

• Laboratory Synthesis: Execute synthesis based on CASP-suggested routes using only available in-house building blocks. Implement high-throughput synthetic workflows where appropriate, such as slurry-based solid-state synthesis for inorganic materials or automated parallel synthesis for organic compounds [35].

• Biological Testing: Evaluate synthesized compounds for biological activity against the target of interest. Promising confirmed hits can inform further iterations of the design-synthesis-test cycle.

Implementation Considerations

• Computational Requirements: The integrated workflow requires significant computational resources, particularly for the virtual screening stage. A local HPC cluster with thousands of CPUs and GPUs is recommended for screening billion-compound libraries within practical timeframes [34].
• Building Block Management: Maintain a digital inventory of available building blocks with accurate structural information. This inventory serves as the foundation for both synthesizability scoring and CASP route generation.
• Model Retraining: Periodically retrain the synthesizability score model to reflect changes in the building block inventory and incorporate new synthesis data from successful and failed attempts.

The Scientist's Toolkit

Table 3: Essential Research Reagents and Computational Tools

Tool/Resource	Type	Function	Example Sources/Platforms
Building Block Collection	Chemical Reagents	Provides foundational compounds for synthesis	In-house collections (e.g., ~6,000 compounds) [27]
Virtual Screening Platform	Software	Identifies potential bioactive compounds from large libraries	OpenVS, RosettaVS [34]
Synthesis Planning Tool	Software	Generates potential synthesis routes for target molecules	AiZynthFinder [27]
Synthesizability Score	Computational Model	Predicts likelihood of successful synthesis	CASP-based or heuristic scoring functions [27]
High-Throughput Synthesis Workflow	Laboratory Equipment	Enables parallel synthesis of multiple candidates	Automated liquid handlers, isopresses [35]
Solid-State Synthesis Materials	Chemical Reagents	Raw materials for inorganic material synthesis	Oxides, carbonates, oxalates [35]
IGF-1R inhibitor-3	IGF-1R inhibitor-3, MF:C28H27FN6O, MW:482.6 g/mol	Chemical Reagent	Bench Chemicals
Isomorellinol	Isomorellinol, MF:C33H38O7, MW:546.6 g/mol	Chemical Reagent	Bench Chemicals

Case Study: MGLL Inhibitor Discovery

A practical implementation of this workflow demonstrated the discovery of novel monoglyceride lipase (MGLL) inhibitors [27]. Researchers first trained an in-house synthesizability score on their available building block collection of approximately 6,000 compounds. They then performed multi-objective de novo drug design optimizing both predicted activity (using a QSAR model) and in-house synthesizability.

The approach generated thousands of potentially active and easily synthesizable candidate molecules. From these, three de novo candidates were selected for experimental evaluation using their CASP-suggested synthesis routes employing only in-house building blocks. The study identified one candidate with evident biochemical activity, demonstrating the practical utility of incorporating in-house synthesizability assessment into the virtual screening workflow [27].

Integrating synthesizability checks into high-throughput virtual screening represents a crucial advancement in computational drug discovery and materials science. By bridging the gap between computational prediction and practical synthesis constraints, this integrated approach significantly increases the efficiency of the discovery pipeline. The methodologies and protocols outlined in this application note provide researchers with a framework for implementing these integrated workflows, potentially accelerating the translation of computational hits into synthesized and tested candidates.

Overcoming Practical Hurdles: Data, Generalization, and Model Interpretation

Addressing Data Scarcity and the Positive-Unlabeled Learning Paradigm

The discovery of new functional materials and molecules is fundamental to technological progress, yet it remains constrained by a fundamental challenge: determining whether a computationally designed compound can be successfully synthesized in the laboratory. This problem of synthesizability prediction is notoriously difficult because, while databases contain records of successfully synthesized "positive" examples, confirmed "negative" examples (compounds that definitively cannot be made) are exceptionally scarce; failed synthesis attempts are rarely published or systematically recorded [1] [36]. This lack of negative data renders standard supervised machine learning approaches suboptimal.

The Positive-Unlabeled (PU) Learning paradigm directly addresses this data scarcity. PU learning is a semi-supervised framework designed to train classification models using only a set of confirmed positive examples and a large set of unlabeled data, the latter of which contains an unknown mixture of positive and negative instances [37]. This approach is particularly powerful for materials and drug discovery, where it can develop an intuition for synthesizability by learning from the patterns embedded within known synthesized materials [38]. When combined with advanced material representations like atom2vec—which learns meaningful vector representations of atoms from a large database of known compounds—PU learning provides a robust and computationally efficient path for identifying promising, synthesizable candidates [1] [5].

Quantitative Performance of PU Learning Methods

PU learning models have demonstrated superior performance compared to traditional heuristic and computational methods for predicting synthesizability. The tables below summarize the quantitative performance of various approaches as reported in recent literature.

Table 1: Performance Comparison of Synthesizability Prediction Methods

Method	Key Principle	Reported Performance	Reference / Model
SynthNN	Deep learning on known compositions; uses atom2vec.	7x higher precision than DFT formation energy; 1.5x higher precision than best human expert.	[1]
PU Learning (General)	Semi-supervised learning from positive/unlabeled data.	True Positive Rate of 0.91 on Materials Project database.	[38]
Charge-Balancing Heuristic	Filters compositions with net neutral ionic charge.	Only 37% of known synthesized ICSD materials are charge-balanced.	[1]
Human Experts	Specialized knowledge and intuition.	Baseline for comparison; outperformed by SynthNN in precision and speed.	[1]

Table 2: Performance of Specific PU Learning Frameworks

Framework	Architecture	Material Class	Key Result
SynCoTrain	Dual GCNN co-training (ALIGNN & SchNet).	Oxide Crystals	Achieved high recall on internal and leave-out test sets.	[36]
Solid-State PU Model	Positive-unlabeled learning on literature data.	Ternary Oxides	Predicted 134 of 4312 hypothetical compositions as synthesizable.	[39]
MXene-Specific PU Model	Decision tree classifier with bootstrapping.	2D MXenes	Identified 18 new MXenes predicted to be synthesizable.	[38]

Detailed Experimental Protocols

This section provides detailed, actionable protocols for implementing two distinct PU learning approaches for synthesizability prediction, one based on compositional data and another on crystal structure.

Protocol 1: Composition-Based Synthesizability Prediction with SynthNN

This protocol is adapted from the SynthNN model, which predicts synthesizability from chemical formulas without requiring structural information [1].

1. Research Reagents and Data Sources

• Primary Data: Curate a list of synthesized inorganic crystalline materials. The Inorganic Crystal Structure Database (ICSD) is the standard source [1].
• Material Representations: Utilize the atom2vec framework to convert chemical formulas into a learned atom embedding matrix, which serves as the optimal input representation [1] [5].
• Software: Standard deep learning libraries (e.g., PyTorch, TensorFlow) are required for implementing the neural network.

2. Step-by-Step Procedure 1. Data Preparation: - Extract all chemical formulas from the ICSD. These are your positive (P) examples. - Generate a large set of artificial chemical formulas that are not present in the ICSD. This set constitutes your unlabeled (U) data. The ratio of artificial to synthesized formulas (e.g., ( N_{synth} )) is a key hyperparameter. 2. Model Architecture Definition: - Design a neural network where the first layer is an embedding layer that learns vector representations for each element in the periodic table (atom2vec). - Follow the embedding layer with fully connected hidden layers and a final output layer with a sigmoid activation function for binary classification. 3. PU Learning Training Loop: - Implement a semi-supervised loss function that treats unlabeled examples as probabilistically weighted negatives. This accounts for the fact that the unlabeled set contains synthesizable materials that have simply not been discovered or reported yet [1]. - Train the model to distinguish the positive examples from the unlabeled set. 4. Validation and Prediction: - Validate model performance by measuring its ability to classify a held-out test set of known ICSD materials versus artificially generated formulas. - Use the trained model to screen a vast space of hypothetical compositions, ranking them by their predicted synthesizability score.

Protocol 2: Structure-Based Synthesizability Prediction with SynCoTrain

This protocol outlines the co-training framework using graph neural networks for materials where crystal structure is available or can be reliably predicted [36].

1. Research Reagents and Data Sources

• Primary Data: Obtain crystal structures from databases like the Materials Project. For oxides, filter data using tools like pymatgen's get_valences function to ensure correct oxidation states [36].
• Software & Models: The pumml Python library provides implementations for PU learning. This protocol specifically requires the ALIGNN and SchNetPack model architectures [38] [36].

2. Step-by-Step Procedure 1. Data Curation and Featurization: - Designate experimentally synthesized crystals (from ICSD) as the positive (P) set. - Designate computationally proposed crystals (from Materials Project, marked 'theoretical') as the unlabeled (U) set. - Remove any experimental data with an energy above hull (Ehull) significantly greater than 1eV, as these may be corrupt entries [36]. - Convert all crystal structures into graph representations, where atoms are nodes and bonds are edges. 2. Dual-Classifier Co-Training Setup: - Initialize two different Graph Convolutional Neural Networks (GCNNs): ALIGNN (encodes bonds and angles) and SchNet (uses continuous-filter convolutional layers). - Each network will function as an independent PU learner. 3. Iterative Co-Training Process: - Step 1: Train the first PU learner (e.g., ALIGNN) on the initial P and U sets using the method of Mordelet and Vert [36]. The model will assign a high synthesizability score to some items in U. - Step 2: Select the most confidently predicted positives from the U set according to the first learner and add them to the labeled positive set for the second learner. - Step 3: Train the second PU learner (e.g., SchNet) on this updated, enlarged positive set and the remaining U set. - Step 4: This learner, in turn, identifies its own set of confident positives from U to add to the positive set for the first learner. - Repeat Steps 1-4 for a predefined number of iterations, allowing the two models to collaboratively enrich the labeled positive data from the unlabeled pool. 4. Prediction and Model Averaging: - After the final co-training iteration, the predicted synthesizability score for a new material is the average of the scores output by the two trained classifiers [36].

The following workflow diagram illustrates the synergistic co-training process between the two classifiers.

The Scientist's Toolkit: Essential Research Reagents

The table below catalogs key computational tools, data sources, and models required for implementing PU learning in synthesizability research.

Table 3: Key Research Reagents for PU Learning in Synthesizability Prediction

Name	Type	Function in Research	Relevance to atom2vec/PU
Inorganic Crystal Structure Database (ICSD)	Data Source	Provides canonical set of synthesized inorganic crystals as positive examples.	Foundational for creating the positive (P) set [1] [36].
Materials Project Database	Data Source	Source of theoretical/unlabeled crystal structures and properties for screening.	Provides the unlabeled (U) set and data for featurization [38] [1].
atom2vec	Material Representation	Learns element embeddings from data; provides model input from composition.	Core representation that learns chemistry from data distribution [1] [5].
ALIGNN & SchNet	Graph Neural Network Models	Encode crystal structure graphs for structure-based prediction.	Used as complementary classifiers in co-training frameworks like SynCoTrain [36].
pumml Library	Software	Python implementation of PU learning for materials science.	Provides pre-built tools for running synthesizability predictions [38].
AiZynthFinder	Software Tool	Computer-Aided Synthesis Planning (CASP) for validating synthesis routes.	Used to generate data for and validate CASP-based synthesizability scores [27].
Hdac-IN-59	Hdac-IN-59, MF:C20H25NO7, MW:391.4 g/mol	Chemical Reagent	Bench Chemicals
Ravenelin	Ravenelin, CAS:479-44-7, MF:C14H10O5, MW:258.23 g/mol	Chemical Reagent	Bench Chemicals

Workflow for Integrating atom2vec and PU Learning

The conceptual workflow below illustrates how atom2vec representation and PU learning integrate into a cohesive materials discovery pipeline, from data curation to experimental validation.

Ensuring Model Generalization Beyond Training Set Compositions

The application of machine learning, particularly deep learning models utilizing atomistic representations, has revolutionized the field of materials discovery by enabling rapid prediction of material properties and synthesizability. A fundamental challenge in this domain lies in ensuring that models generalize effectively beyond the specific compositions present in their training data. Models that fail to generalize cannot reliably predict the synthesizability of novel, unexplored chemical compositions, severely limiting their utility for genuine materials discovery. The Atom2Vec framework and its derivatives offer a promising approach to this challenge by learning distributed representations of atoms from extensive materials databases, capturing fundamental chemical properties that transfer across compositional space [5] [4]. This application note details protocols and methodologies for developing and validating synthesizability prediction models with robust generalization capabilities, framed within the broader context of chemical representation research.

Foundational Concepts and Representation Learning

Atomistic Representation Frameworks

Atom2Vec serves as a foundational unsupervised learning approach that generates distributed representations of atoms by analyzing the co-occurrence patterns of elements in known crystal structures [5]. Inspired by natural language processing techniques like Word2Vec, Atom2Vec processes crystal structures as "sentences" where atoms are "words," learning vector representations that encapsulate chemical similarities. Remarkably, this approach can reconstruct fundamental chemical groupings from the periodic table without explicit supervision, demonstrating its capacity to learn chemically meaningful representations [5] [40].

SkipAtom represents an evolution of this concept, employing a skip-gram model that predicts context atoms given a target atom within materials' crystal structures [4]. By representing materials as graphs derived from Voronoi decomposition of crystal structures, SkipAtom learns atomic embeddings that reflect chemo-structural environments. The model maximizes the average log probability expressed as:

$$\frac{1}{| M| }\mathop{\sum}\limits{m\in M}\mathop{\sum}\limits{a\in {A}{m}}\mathop{\sum}\limits{n\in N(a)}\log p(n| a)$$

where (M) represents the set of materials, (A_{m}) the atoms in material (m), and (N(a)) the neighbors of atom (a) [4]. This approach effectively captures the complex relationships between atoms and their local environments, creating representations that facilitate generalization.

The Positive-Unlabeled Learning Paradigm

A significant challenge in synthesizability prediction is the absence of definitively labeled negative examples (unsynthesizable materials) in materials databases. Positive-Unlabeled (PU) learning addresses this issue by treating unlabeled materials as probabilistically weighted examples rather than definitive negatives [1] [21]. This approach acknowledges that while synthesized materials represent positive examples, the absence of a material from databases does not definitively indicate unsynthesizability. In synthesizability prediction, PU learning frameworks like those used in SynthNN treat artificially generated formulas as unlabeled data, reweighting them according to their likelihood of being synthesizable [1]. This methodology more accurately reflects the real-world scenario where most potential materials remain unexplored, enhancing model generalization by preventing overconfident negative classifications.

Quantitative Performance Comparison

Table 1: Comparative Performance of Synthesizability Prediction Methods

Method	Key Principle	Accuracy/Precision	Generalization Strengths
SynthNN	Deep learning on entire space of synthesized compositions	7× higher precision than DFT formation energies [1]	Learns charge-balancing, chemical family relationships, and ionicity without prior knowledge [1]
Charge-Balancing	Net neutral ionic charge according to common oxidation states	Only 37% of known materials charge-balanced [1]	Limited generalization due to inflexible constraint [1]
CSLLM	Large language model fine-tuned on crystal structures	98.6% accuracy [22]	Exceptional generalization to complex structures with large unit cells [22]
PU Learning (Solid-State)	Positive-unlabeled learning from human-curated data	Effective identification of synthesizable ternary oxides [21]	Addresses data sparsity and labeling uncertainty [21]

Table 2: Atomistic Representation Methods for Generalization

Representation Method	Input Data	Learning Approach	Generalization Capabilities
Atom2Vec	Co-occurrence of atoms in known compounds	Unsupervised from materials database [5]	Captures periodic table relationships and chemical similarities [5]
SkipAtom	Local atomic connectivity in crystal structures	Skip-gram model on material graphs [4]	Learns chemo-structural relationships; enables composition-based property prediction competitive with structure-based methods [4]
Mat2Vec	Scientific text and abstracts	Word2Vec on materials science literature [4]	Captures contextual relationships from research literature [4]
ATM (AtomTransMachine)	Spatial relationships in molecular structures	Self-supervised with multi-attention mechanism [41]	Decouples mutual features between atoms; captures similarities among family or adjacent elements [41]

Experimental Protocols

Protocol: Implementing Cross-Domain Validation

Purpose: To evaluate model generalization across distinct chemical families not represented in training data.

Procedure:

• Partition Dataset by Chemical Families: Divide materials database into distinct chemical families based on composition (e.g., oxides, sulfides, intermetallics, perovskites).
• Train-Test Split Strategy: Implement leave-one-family-out cross-validation where models are trained on all but one chemical family and tested on the excluded family.
• Model Training:
  • Utilize Atom2Vec or SkipAtom to generate atomic embeddings [5] [4].
  • Implement SynthNN architecture with learned embeddings as input features [1].
  • Train with PU learning framework to handle unlabeled examples [1] [21].
• Generalization Metrics:
  • Calculate precision/recall on excluded chemical families.
  • Compare performance against random guessing and charge-balancing baselines.
  • Analyze performance degradation relative to within-family prediction.

Interpretation: Models demonstrating less than 30% performance degradation on excluded chemical families exhibit acceptable generalization. Significant drops indicate overfitting to training compositions.

Protocol: Data-Efficient Grammar Learning for Molecular Generation

Purpose: To generate synthesizable molecules with limited training data through learnable graph grammar.

Procedure:

• Graph Representation: Represent molecular structures as graphs with atoms as nodes and chemical bonds as edges [42].
• Grammar Induction:
  • Identify recurring substructures (two atoms connected by a bond, short sequences of bonded atoms, rings of atoms) in training molecules.
  • Collapse substructures to single nodes repeatedly to create production rules.
  • Define a minimal collection of production rules that maximize synthesizability of generated molecules [42].
• Model Training:
  • Apply DEG (Data-Efficient Graph Grammar) to learn optimal production rules from limited training data (<50 samples) [42].
  • Optimize grammar to maximize percentage of generated molecules that are synthesizable.
• Validation:
  • Generate new molecules by applying production rules in novel combinations.
  • Evaluate synthesizability of generated molecules using SynthNN or CSLLM [1] [22].
  • Compare against state-of-the-art models for percentage of chemically valid and unique molecules.

Interpretation: Successful generalization is indicated by generation of >80% chemically valid molecules with <50 training samples, outperforming conventional deep learning approaches that require thousands of examples [42].

Protocol: Composition-Based Property Prediction

Purpose: To predict material properties using only compositional information without structural data.

Procedure:

• Representation Generation:
  • Employ SkipAtom or Atom2Vec to generate atomic embeddings [5] [4].
  • For compound representations, apply pooling operations (e.g., mean, sum, weighted average) to constituent atom vectors [4].
• Model Architecture:
  • Implement neural network with embedding layer, multiple hidden layers, and output layer suitable for prediction task (classification for synthesizability, regression for formation energy).
  • Use compositional embeddings as input rather than traditional descriptor sets.
• Training with Limited Data:
  • Utilize semi-supervised learning approaches that combine limited labeled data with extensive unlabeled materials data [4].
  • Apply transfer learning from related property predictions when available.
• Evaluation:
  • Assess performance on diverse composition spaces excluding structurally similar materials from training.
  • Compare against structure-based models to establish performance gap.

Interpretation: Composition-based models achieving >80% of structure-based model performance demonstrate effective generalization, particularly valuable for screening novel compositions where structure is unknown [4].

Visualization of Workflows

Synthesizability Prediction Pipeline

Synthesizability Prediction Workflow: This diagram illustrates the complete pipeline for predicting synthesizability of novel compositions, from materials database to final prediction, highlighting the role of atomic representation learning and PU learning frameworks.

Atomistic Representation Learning

Atomistic Representation Learning: This workflow visualizes the process of learning distributed atomic representations from crystal structures, capturing chemical similarities that enable generalization.

The Scientist's Toolkit: Essential Research Reagents

Table 3: Key Computational Tools for Generalization Research

Tool/Resource	Type	Function in Generalization Research	Access
Inorganic Crystal Structure Database (ICSD)	Materials Database	Primary source of synthesizable materials for training; provides positive examples for PU learning [1] [21]	Commercial
Materials Project	Computational Database	Source of hypothetical structures as unlabeled examples; provides formation energies for benchmarking [21]	Free
Atom2Vec/SkipAtom	Representation Learning	Generates transferable atomic embeddings that capture chemical similarities [5] [4]	Open Source
SynthNN	Deep Learning Model	Implements synthesizability classification with PU learning framework [1]	Research Code
DEG (Data-Efficient Graph Grammar)	Molecular Generation	Enables generation of synthesizable molecules with limited training data [42]	Research Code
CAF/SAF (Composition/Structure Analyzer Featurizer)	Feature Generation	Provides explainable compositional and structural features for interpretable ML [3]	Open Source
LC3in-C42	LC3in-C42, MF:C19H21Cl2N3O2, MW:394.3 g/mol	Chemical Reagent	Bench Chemicals
Crocacin A	Crocacin A, CAS:157698-34-5, MF:C31H42N2O6, MW:538.7 g/mol	Chemical Reagent	Bench Chemicals

Ensuring model generalization beyond training set compositions remains a critical challenge in computational materials discovery. The integration of atomistic representation learning with Positive-Unlabeled frameworks provides a robust foundation for developing synthesizability models that transfer effectively to novel chemical spaces. As demonstrated by SynthNN, CSLLM, and grammar-based approaches, these methodologies can outperform traditional stability metrics and even human experts in predicting synthesizability of unexplored compositions [1] [22]. Future research directions include developing dynamic representation learning that adapts to new synthetic capabilities, integrating multi-modal data from synthesis recipes and conditions, and creating increasingly data-efficient models that minimize requirements for expensive experimental data. By adopting the protocols and methodologies outlined in this application note, researchers can develop predictive models with enhanced generalization capabilities, accelerating the discovery of novel functional materials.

The adoption of artificial intelligence (AI) in chemical and materials science has revolutionized the discovery of new compounds. However, many advanced machine learning models operate as "black boxes," providing accurate predictions without revealing their underlying reasoning [43]. This opacity is a significant barrier in scientific fields, where understanding the why behind a prediction is as crucial as the prediction itself [44]. The nascent field of Explainable AI (XAI) seeks to make these models more transparent and interpretable [43].

This Application Note focuses on interpreting black-box models within the specific context of chemical synthesizability research. We detail how models, particularly those using atom2vec-inspired representations, can be probed to uncover the fundamental chemical rules they infer from data. The protocols herein are designed for researchers and scientists aiming to validate model predictions and extract new, actionable chemical insights.

Background: atom2vec and Material Representations

A pivotal step in applying machine learning to chemistry is representing atoms and compounds in a form digestible by algorithms. The atom2vec approach aims to derive distributed representations of atoms, where each atom is represented by a vector in a continuous space [5] [4].

• Core Principle: Inspired by natural language processing, the hypothesis is that "atoms, like words, can be understood by the company they keep" [4]. An atom's vector representation is learned from its co-occurrence with other atoms in known crystal structures or chemical databases.
• Advantages over Traditional Representations: Unlike simple one-hot vectors or random vector assignments, learned embeddings capture semantic relationships between atoms [4]. Atoms with similar chemical properties or roles are positioned closer together in the vector space, allowing the model to leverage chemical intuition during learning.
• From Atoms to Materials: Representations of entire chemical formulas are constructed by pooling the vectors of their constituent atoms. This allows for the representation of materials even when their full crystal structure is unknown, enabling high-throughput screening of compositional space [4].

Quantitative Evidence of Learned Chemical Rules

Evidence suggests that models utilizing these learned representations internalize complex chemical principles without being explicitly programmed with them. The following table summarizes key findings from a synthesizability prediction model (SynthNN) that leverages the entire space of synthesized inorganic compositions [1].

Table 1: Quantitative Evidence of Chemical Rules Learned by the SynthNN Model

Learned Chemical Principle	Experimental Evidence from Model Analysis	Performance Implication
Charge-Balancing	The model learned to prioritize compositions with net neutral ionic charge, despite only 37% of known synthesized inorganic materials being perfectly charge-balanced according to common oxidation states [1].	Identifies synthesizable materials with 7x higher precision than DFT-calculated formation energies alone [1].
Chemical Family Relationships	The model successfully clusters and treats atoms from the same chemical family (e.g., alkali metals, halogens) similarly, based on their co-occurrence in known structures [1] [4].	Achieves 1.5x higher precision than the best human expert in a head-to-head material discovery challenge [1].
Ionicity	Analysis of binary cesium compounds showed the model learned nuances of ionic bonding beyond a simple charge-neutrality filter, which only applies to 23% of known ionic binaries [1].	Completes synthesizability screening tasks five orders of magnitude faster than human experts [1].

Experimental Protocols for Interpreting Black-Box Chemical Models

This section provides a detailed methodology for probing a black-box model to uncover the chemical rules it has learned. The workflow assumes a model trained to predict a chemical property (e.g., synthesizability, formation energy) using atom2vec-style compositional embeddings.

Protocol 1: Model Interpretation via Layer-wise Relevance Propagation (LRP)

Purpose: To decompose a model's prediction for a specific compound into contributions from its n-body atomic interactions, revealing which subsets of atoms most strongly influenced the output.

Materials:

• Trained Black-Box Model: A neural network potential (NNP) or graph neural network (GNN) for material property prediction.
• Input Data: The chemical composition and/or crystal structure of the material to be interpreted.
• Software: GNN-LRP tools (e.g., extensions of the methods described in [45]).

Procedure:

• Model Forward Pass: Input the target material's representation into the trained model to obtain the prediction (e.g., synthesizability score).
• Relevance Propagation: Apply the LRP algorithm backward through the network. LRP operates by decomposing the activation of each neuron into contributions from its inputs, layer by layer, until the input features are reached [45].
• Attribute Relevance to Input Features: The output of LRP is a set of relevance scores assigned to the input features. For a GNN, this involves attributing relevance to sequences of graph edges ("walks") that connect atoms in the input structure [45].
• Aggregate n-body Contributions: Group the relevance scores of all "walks" associated with a particular subgraph (e.g., a pair or triplet of atoms) to determine the total relevance of that n-body interaction to the final prediction.
• Analysis: Identify the atomic interactions (e.g., specific cation-anion pairs, three-body angle constraints) with the highest absolute relevance scores. These are the interactions the model deemed most critical for its prediction.

Protocol 2: Validating Learned Rules via Ablation and Dimensional Analysis

Purpose: To test the hypothesis that a model relies on specific chemical principles by systematically perturbing its inputs and analyzing the effects on its predictions.

Materials:

• Trained Black-Box Model (as in Protocol 1).
• Validation Dataset: A curated set of chemical compositions with known properties.

Procedure:

• Establish Baseline: Run the model on the unmodified validation dataset and record its prediction accuracy.
• Define Perturbation Strategy: Design input perturbations that violate a specific chemical rule. For example:
  • Charge Imbalance: Systematically alter compositions to create a net ionic charge.
  • Element Substitution: Replace an atom with another from a different chemical group while keeping the overall structure similar.
• Perform Ablation Study: Execute the model on the perturbed dataset.
• Analyze Performance Shift: Compare the model's performance on the perturbed data against the baseline. A significant drop in accuracy strongly indicates that the model had learned and depended on the ablated rule for its predictions on the original data [1].
• Dimensional Analysis: Use dimensionality reduction techniques (e.g., t-SNE, PCA) on the model's internal atom embeddings. Visualize the low-dimensional projection. The emergence of clusters corresponding to chemical groups (e.g., all halogens clustered together) provides direct evidence that the model has learned these fundamental relationships [4].

Workflow Visualization

The following diagram illustrates the logical flow of the interpretation process, from model training to scientific insight.

The Scientist's Toolkit: Essential Reagents & Materials

Table 2: Key "Research Reagent Solutions" for Interpreting Chemical ML Models

Item Name	Function & Application
Inorganic Crystal Structure Database (ICSD)	A comprehensive database of known crystalline inorganic structures. Serves as the primary source of "synthesized" materials for training and benchmarking synthesizability models [1].
atom2vec / SkipAtom Embeddings	Pre-trained or custom-learned distributed vector representations of atoms. These serve as the foundational input features for composition-based models, encoding chemical similarity [4].
Graph Neural Network (GNN) Architecture	A type of neural network that operates directly on graph-structured data. Ideal for representing crystal structures where atoms are nodes and bonds are edges, enabling the model to learn from local chemical environments [45].
GNN-LRP Software Tools	Implementation of Layer-wise Relevance Propagation for Graph Neural Networks. The core software for executing Protocol 1, allowing for the decomposition of model predictions into n-body interaction contributions [45].
Dimensionality Reduction Suite (e.g., t-SNE, UMAP)	Software tools for projecting high-dimensional atom embeddings into 2D or 3D space. Used in Protocol 2 to visually validate that the model has learned meaningful chemical groupings [4].
2-Chloronaphthalene	2-Chloronaphthalene, CAS:51569-12-1, MF:C10H7Cl, MW:162.61 g/mol
Cdk7-IN-21	Cdk7-IN-21, CAS:2766124-39-2, MF:C33H36FN9O2, MW:609.7 g/mol

In the field of computational materials science, the discovery of new functional materials often hinges on accurately predicting the synthesizability of inorganic crystalline compounds. The atom2vec representation, an unsupervised machine learning algorithm that learns distributed vector representations of atoms from known chemical compounds, has emerged as a powerful tool for this task [5] [4]. By reformulating material discovery as a synthesizability classification task, models like SynthNN leverage these representations to identify synthesizable materials with significantly higher precision than traditional methods such as density-functional theory (DFT)-calculated formation energies [1]. However, a central challenge persists: the trade-off between the computational cost of generating predictions and the prediction accuracy of the models. This application note provides a detailed framework for managing this trade-off within the context of atom2vec-enabled synthesizability research, offering structured data, experimental protocols, and practical toolkits for researchers.

Quantitative Landscape of Cost vs. Accuracy

Selecting an appropriate model and representation requires a clear understanding of their performance and computational demands. The table below summarizes key metrics for different approaches to synthesizability prediction, highlighting the distinct advantages of representation learning methods like atom2vec.

Table 1: Comparison of Synthesizability Prediction Methods

Method	Key Input	Reported Accuracy/Precision	Key Advantage	Computational Cost
SynthNN (atom2vec) [1]	Chemical Composition	7x higher precision than DFT formation energy	High throughput; no crystal structure required	Low (composition-based)
CSLLM (Synthesizability LLM) [22]	Crystal Structure (Text Representation)	98.6% Accuracy	State-of-the-art accuracy; suggests synthesis routes	High (structure-based, large model)
DFT (Formation Energy) [1] [22]	Crystal Structure	~74.1% Accuracy (as a synthesizability proxy)	Strong theoretical foundation	Very High (ab-initio calculation)
Charge-Balancing [1]	Chemical Composition	Low (Only 37% of known materials are charge-balanced)	Extremely fast and simple	Very Low (rule-based)

The data reveals a clear hierarchy. While DFT calculations and advanced models like CSLLM can achieve high accuracy, they incur significant computational costs, either from the ab-initio calculations themselves or from the need for detailed crystal structure information [22]. In contrast, the atom2vec-based SynthNN model offers a favorable balance, achieving high precision by learning optimal descriptors directly from composition data, thus bypassing the need for expensive structural simulations [1].

Optimization Strategies and Experimental Protocols

Effectively balancing cost and accuracy involves strategic decisions at both the data representation and model training levels. The following protocol outlines a methodology for developing a cost-effective synthesizability prediction pipeline.

Protocol: A Multi-Fidelity Workflow for Synthesizability Screening

Principle: Leverage a cascade of models with increasing complexity and cost to screen large chemical spaces, reserving the most expensive resources only for the most promising candidates [46].

Procedure:

• Primary Screening with Composition-Based Models

  • Input: A large dataset of candidate chemical formulas.
  • Action: Utilize a pre-trained atom2vec model (e.g., SynthNN) to generate distributed representations for each formula and perform an initial synthesizability classification [1] [4].
  • Output: A subset of candidate materials predicted to be synthesizable. This step rapidly reduces the search space with minimal computational expense.

• Secondary Validation with Structural Models

  • Input: The subset of promising candidates from Step 1.
  • Action: For each candidate, obtain or generate a plausible crystal structure. Subsequently, apply a more accurate, structure-based model like CSLLM for a higher-fidelity synthesizability assessment [22].
  • Output: A refined list of high-confidence, synthesizable materials.

• Final Energetic Validation (Optional)

  • Input: The final, refined list of candidates from Step 2.
  • Action: Perform DFT calculations to verify thermodynamic stability via formation energy or energy above the convex hull [22].
  • Output: A shortlist of synthesizable and thermodynamically viable materials for experimental pursuit.

This multi-fidelity approach ensures that computationally intensive methods are deployed sparingly, optimizing the overall cost-accuracy trade-off [46].

Strategy: Model Compression and Efficient Architectures

To further reduce the inference cost of deep learning models like SynthNN, consider the following techniques post-training:

• Pruning: Systematically remove redundant neurons or weights from the neural network that contribute minimally to the final prediction, reducing model size and complexity [47].
• Quantization: Convert the model's parameters from 32-bit floating-point numbers to lower-precision representations (e.g., 16-bit or 8-bit integers). This shrinks the model and accelerates inference on supported hardware with a typically minor impact on accuracy [47].

Diagram: Multi-Fidelity Synthesizability Screening Workflow

The Scientist's Toolkit: Research Reagent Solutions

Building and applying atom2vec models requires a suite of software tools and data resources. The following table details essential components of the research toolkit.

Table 2: Essential Research Tools and Resources

Tool/Resource	Type	Function in Research	Reference
Inorganic Crystal Structure Database (ICSD)	Database	Primary source of positive (synthesized) examples for training models like SynthNN.	[1] [22]
atom2vec / SkipAtom	Algorithm	Generates unsupervised distributed representations of atoms from materials data, serving as the foundational input for models.	[1] [5] [4]
Positive-Unlabeled (PU) Learning	Machine Learning Framework	Addresses the lack of confirmed negative data by treating un-synthesized materials as unlabeled examples, crucial for robust model training.	[1] [22]
Composition Analyzer Featurizer (CAF)	Software Tool	Generates human-interpretable, numerical compositional features from chemical formulas that can be used alongside or compared to atom2vec vectors.	[3]
Crystal Structure Text Representation (e.g., Material String)	Data Preprocessing	Converts crystal structures into a compact text format enabling the use of large language models (LLMs) for high-accuracy synthesizability prediction.	[22]
Cdk7-IN-28	Cdk7-IN-28, MF:C23H27F3N6O, MW:460.5 g/mol	Chemical Reagent	Bench Chemicals

Diagram: atom2Vec's Role in the Synthesizability Prediction Pipeline

The discovery of new functional materials, particularly metastable phases with desirable catalytic, electronic, or magnetic properties, represents a frontier in materials science. Unlike thermodynamically stable phases, metastable materials possess Gibbs free energy higher than the equilibrium state but persist due to kinetic constraints that prevent their transformation to more stable forms [48]. Traditional methods for predicting material synthesizability have relied heavily on thermodynamic stability calculations derived from density functional theory (DFT), which fail to account for the complex kinetic and synthetic factors that enable metastable phase formation [1] [49].

This case study explores the transformative potential of machine learning approaches, specifically atomistic representations learned through methods like Atom2Vec, in predicting the synthesizability of metastable and kinetically stabilized materials. By reframing material discovery as a synthesizability classification task, these data-driven models capture complex chemical relationships that extend beyond traditional charge-balancing or formation energy criteria [1]. Within the broader context of chemical formula synthesizability research, these learned representations demonstrate remarkable capability in identifying synthesizable materials across the vast composition space of inorganic crystalline compounds.

Theoretical Foundation: From Atomic Representations to Synthesizability Predictions

The Challenge of Predicting Metastable Phases

Metastable phases offer significant scientific interest due to their distinct properties, high-energy structures, and unique electronic environments that often outperform their stable counterparts in catalytic applications [48]. However, their thermal instability and the complex thermodynamic-kinetic balance required for their synthesis present substantial challenges for prediction. Traditional thermodynamic phase diagrams provide essential predictive insights for stable phases but fail to account for the complex formation of non-equilibrium products under fluctuating temperature and pressure conditions [48].

The synthesis of metastable materials often occurs through highly non-equilibrium processes in supersaturated media, at ultra-high pressure, or at low temperatures with suppressed species diffusion [49]. These conditions create a multidimensional reaction space where kinetic factors often dominate thermodynamic driving forces, making synthesizability prediction exceptionally challenging with conventional computational approaches.

Learned Atomic Representations

The core innovation in recent synthesizability prediction research involves the application of distributed atomic representations learned from existing materials databases. These representations treat atoms analogously to words in natural language processing, where meaning is derived from contextual relationships [5] [4].

Atom2Vec represents each chemical element as a high-dimensional vector derived by applying unsupervised learning to the extensive database of known compounds and materials [5]. The algorithm generates these representations by creating a co-occurrence count matrix of atoms and their chemical environments from existing materials databases, then applying singular value decomposition to this matrix [4]. This approach enables the model to capture complex chemical relationships without prior knowledge of explicit chemical rules.

Table 1: Comparison of Atomic Representation Methods

Method	Approach	Data Source	Key Advantages
Atom2Vec	Co-occurrence matrix + SVD	Materials databases (e.g., ICSD)	Captures structural chemistry; directly learned from material compositions
Mat2Vec	Word2Vec algorithm	Scientific literature abstracts	Incorporates research context; captures emerging trends
SkipAtom	Skip-gram model	Crystal structure databases	Learns from local atomic environments; accessible to researchers

These distributed representations encode fundamental chemical properties, with vectors for similar atoms clustering together in the learned space. For example, alkali metals naturally group together, as do light non-metals, demonstrating that the model captures periodic trends without explicit supervision [50]. This emergent organization provides a principled structural foundation for predicting chemical behavior and synthesizability.

Quantitative Assessment of Synthesizability Prediction Models

Performance Comparison: SynthNN vs. Traditional Methods

The synthesizability neural network (SynthNN) model, which leverages Atom2Vec representations, demonstrates superior performance compared to traditional synthesizability assessment methods. When evaluated against charge-balancing criteria and random guessing baselines, SynthNN achieves significantly higher precision in identifying synthesizable materials [1].

Table 2: Performance Comparison of Synthesizability Prediction Methods

Method	Precision	Key Principles	Limitations
SynthNN (Atom2Vec)	7× higher than DFT formation energy	Learned from distribution of synthesized materials	Requires large training datasets
Charge-Balancing	Low (23-37% of known compounds)	Net neutral ionic charge	Inflexible; fails for metallic/covalent materials
DFT Formation Energy	Captures only 50% of synthesized materials	Thermodynamic stability	Misses kinetically stabilized phases

Remarkably, in a head-to-head material discovery comparison against 20 expert material scientists, SynthNN outperformed all human experts, achieving 1.5× higher precision while completing the task five orders of magnitude faster than the best human expert [1]. This demonstrates the potential for such models to dramatically accelerate materials discovery while improving prediction accuracy.

Learned Chemical Principles

Without explicit programming of chemical knowledge, SynthNN internalizes fundamental chemical principles through its training on the Inorganic Crystal Structure Database (ICSD). Experimental analyses indicate the model learns the principles of charge-balancing, chemical family relationships, and ionicity, utilizing these to generate synthesizability predictions [1]. This emergent understanding is particularly valuable for predicting metastable phases, where multiple competing factors influence synthetic accessibility.

The model demonstrates particular effectiveness in identifying synthesizable materials that violate simple charge-balancing heuristics. While only 37% of known inorganic materials in the ICSD are charge-balanced according to common oxidation states (dropping to just 23% for binary cesium compounds), SynthNN successfully identifies many of the "exceptional" cases where other bonding factors overcome charge imbalance [1].

Experimental Protocols and Methodologies

SynthNN Model Development Protocol

Objective: To develop a deep learning classification model for predicting synthesizability of inorganic chemical formulas without structural information.

Materials and Data Sources:

• Primary Data: Crystalline inorganic materials from the Inorganic Crystal Structure Database (ICSD)
• Training Framework: Positive-unlabeled (PU) learning approach
• Data Split: Standard train/validation/test partition (e.g., 80/10/10)

Procedure:

• Data Extraction and Preprocessing:
  • Extract chemical formulas of synthesized crystalline inorganic materials from ICSD
  • Apply formula standardization and normalization
  • Generate artificially unsynthesized materials for negative examples

• Representation Learning:

  • Initialize Atom2Vec embeddings for each element
  • Represent chemical formulas as learned atom embedding matrices
  • Optimize embedding dimensionality through hyperparameter tuning

• Model Architecture and Training:

  • Implement deep neural network architecture with embedding layer
  • Apply semi-supervised learning to handle unlabeled examples
  • Probabilistically reweight unsynthesized materials according to synthesizability likelihood
  • Train with cross-entropy loss and Adam optimizer

• Validation and Testing:

  • Evaluate using standard performance metrics (precision, recall, F1-score)
  • Compare against charge-balancing and random guessing baselines
  • Perform ablation studies to assess contribution of different model components

Metastable Phase Synthesis Validation Protocol

Objective: To experimentally validate predicted metastable phases through targeted synthesis.

Materials:

• Precursors: High-purity elemental powders or precursor compounds
• Equipment: Solid-state reaction setup (tube furnaces, quartz ampoules) or solution-based synthesis apparatus

Procedure:

• Synthesis Planning:
  • Select synthesis route based on predicted metastable phase (solid-state, solvothermal, CVD)
  • Determine appropriate temperature, pressure, and atmosphere conditions

• Phase Stabilization:

  • Apply kinetic trapping strategies (rapid quenching, epitaxial stabilization)
  • Utilize substrate effects for thin-film metastable phases
  • Implement templating approaches for structural control

• Characterization:

  • Perform X-ray diffraction for phase identification
  • Conduct electron microscopy for morphological analysis
  • Employ spectroscopic techniques (XPS, Raman) for electronic structure verification

• Property Validation:

  • Measure catalytic activity for target reactions
  • Characterize electronic properties relevant to application
  • Assess thermal stability and phase transition behavior

Workflow Visualization: Atom2Vec for Synthesizability Prediction

Synthesizability Prediction with Atom2Vec Representations

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Research Materials for Metastable Phase Synthesis and Validation

Material/Reagent	Function/Purpose	Application Context
High-Purity Elemental Powders	Precursors for solid-state synthesis	Bulk metastable phase formation
Single-Crystal Substrates	Epitaxial stabilization template	Thin-film metastable phase growth
Solvothermal Media	Low-temperature synthesis environment	Kinetic trapping of metastable phases
Chemical Transport Agents	Facilitate vapor-phase crystal growth	Single-crystal metastable phase preparation
High-Pressure Anvils	Create non-ambient synthesis conditions	High-pressure metastable polymorphs
Rapid Quenching Apparatus	Kinetic trapping of high-temperature phases	Glassy and amorphous materials
Structural Characterization Standards	Reference materials for phase identification	XRD, TEM, and spectroscopy calibration

Applications in Metastable Phase Catalysis

Metastable phase materials exhibit exceptional promise in catalytic applications including photocatalysis, electrocatalysis, and thermal catalysis due to their tunable electronic structures, high-energy configurations, and unique surface properties [48]. The strong interaction between metastable phases and reactant molecules, attributed to their easily tunable d-band center and high Gibbs free energy, enables optimization of reaction barriers and accelerated kinetics [48].

The integration of Atom2Vec-derived synthesizability predictions with metastable phase catalysis research creates a powerful feedback loop. Successful synthesis of predicted metastable catalysts validates the prediction approach, while the enhanced catalytic performance of these materials provides functional motivation for continued synthesizability research. This synergy is particularly valuable for identifying novel catalytic materials that might be overlooked by traditional thermodynamic screening approaches.

The application of Atom2Vec representations for predicting synthesizability of metastable and kinetically stabilized materials represents a paradigm shift in computational materials discovery. By learning chemical principles directly from experimental data rather than relying on predefined descriptors, these models capture the complex interplay of factors that influence synthetic accessibility. The demonstrated superiority of SynthNN over both traditional computational methods and human experts highlights the transformative potential of this approach.

Future developments in this field will likely focus on integrating time-dependent synthetic parameter predictions with composition-based synthesizability assessments, enabling not just identification of synthesizable materials but also guidance on optimal synthesis conditions. Additionally, the extension of these approaches to dynamic metastability, where materials properties evolve under operational conditions, presents an exciting frontier for functional materials design. As materials databases continue to grow and representation learning methods become more sophisticated, the accuracy and scope of synthesizability predictions for metastable phases will continue to improve, accelerating the discovery of next-generation functional materials.

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

Application Note

This application note details the transformative potential of the deep learning synthesizability model, SynthNN, which leverages the atom2vec representation to predict the synthesizability of inorganic crystalline materials. Benchmarked against traditional density functional theory (DFT)-based formation energy calculations and human expert intuition, SynthNN demonstrates a paradigm shift in the accuracy and efficiency of virtual materials screening. By reformulating material discovery as a synthesizability classification task, SynthNN achieves a 7-fold higher precision than formation energy-based approaches and outperforms the best human expert by 1.5-fold in precision while completing the task five orders of magnitude faster [1] [51]. This protocol provides a comprehensive guide to implementing SynthNN and its underlying atom2vec featurization to enhance the reliability of computational materials discovery pipelines.

The discovery of novel functional materials is a cornerstone of technological advancement. However, the initial critical step—identifying a novel chemical composition that is synthesizable—remains a significant bottleneck [1]. Traditional computational screening methods have relied heavily on DFT-calculated formation energies as a proxy for stability and synthesizability. While materials with negative formation energies are thermodynamically stable, this metric alone is an imperfect predictor of synthetic accessibility. Numerous metastable materials are synthetically accessible, while many theoretically stable compounds remain elusive [52] [22]. Furthermore, human experts, though invaluable, are limited by their domain-specific knowledge and the sheer vastness of the chemical space. The development of SynthNN addresses these limitations by learning the complex, implicit "rules" of synthesizability directly from the entire history of synthesized inorganic materials, thereby providing a data-driven solution to a historically intuition-driven problem [1].

Quantitative Performance Breakdown

The performance of SynthNN was rigorously quantified against established baselines and human experts. The key results are summarized in the table below.

Table 1: Performance comparison of synthesizability assessment methods

Method	Key Metric	Performance	Inference Time
DFT Formation Energy	Precision in identifying synthesizable materials	Baseline (1x)	Hours to days (for single calculation)
Charge-Balancing Heuristic	Precision in identifying synthesizable materials	Lower than SynthNN [1]	Seconds
Human Experts (Best)	Precision in discovery task	1.5x lower than SynthNN [1]	Days to weeks
SynthNN (atom2vec)	Precision in identifying synthesizable materials	7x higher than DFT [1] [51]	Seconds (for millions of compositions) [1]

Beyond raw precision, a key advantage of SynthNN is its ability to learn complex chemical principles without explicit programming. Experiments indicate that the model internally learns the importance of charge-balancing, chemical family relationships, and ionicity, moving beyond the rigid and often inaccurate charge-balancing heuristic that only applies to ~37% of known synthesized materials [1].

The Scientist's Toolkit: Essential Research Reagents

The following table lists the key computational "reagents" required to implement and utilize the SynthNN framework.

Table 2: Key research reagents and resources for SynthNN implementation

Item	Function / Description	Source / Example
Inorganic Crystal Structure Database (ICSD)	Primary source of positive training data; a collection of experimentally synthesized and characterized inorganic crystal structures [1].	FIZ Karlsruhe
atom2vec Representation	A learned, dense vector representation for each element that captures chemical similarity from the data distribution; serves as the input feature for SynthNN [1] [3].	Custom implementation from model training
Positive-Unlabeled (PU) Learning Algorithm	A semi-supervised learning framework that treats non-synthesized materials as unlabeled data, accounting for the lack of definitive negative examples [1] [22].	Integrated into SynthNN training
Deep Learning Framework	Software environment for constructing, training, and deploying deep neural networks (e.g., TensorFlow, PyTorch).	Open-source
Materials Screening Pipeline	Computational workflow for generating and screening candidate material compositions.	e.g., Materials Project [52]

Experimental Protocols

Protocol 1: Data Curation and theatom2vecFeaturization

Objective: To prepare a dataset of chemical compositions and convert them into a numerical format suitable for training the SynthNN model.

Background: The atom2vec framework learns a distributed representation for each chemical element by leveraging the co-occurrence statistics of elements in known synthesized materials. This method is analogous to word embedding techniques in natural language processing, where the model learns that elements appearing in similar chemical contexts have similar vector representations [1] [3]. This learned representation is more expressive than using fixed, hand-engineered elemental properties.

Materials:

• Inorganic Crystal Structure Database (ICSD) [1]
• Computational resources for data processing (e.g., Python, Pandas)

Procedure:

• Data Extraction: Download the entire ICSD or a filtered subset containing only crystalline inorganic materials. Extract the chemical formulas for all entries.
• Formula Parsing and Normalization: Parse each chemical formula into its constituent elements and their stoichiometric ratios. Normalize formulas to a standard format (e.g., sorted by electronegativity [3]).
• Generate atom2vec Embeddings: a. Represent each chemical formula as a sequence of elements, weighted by their stoichiometric coefficients or as a bag-of-atoms. b. Train an embedding model (e.g., a skip-gram model) on the entire corpus of ICSD formulas. The objective is to predict context elements given a target element. c. Treat each element as a unique token and learn a dense vector (e.g., 50-100 dimensions) for each. The dimensionality is a key hyperparameter. d. The final trained model contains the atom2vec embedding matrix, where each row is the vector representation of an element.
• Featurize Compositions: For a given chemical formula (e.g., NaCl), convert it into an input vector by combining the atom2vec vectors of its constituent elements (Na and Cl). This can be done through a weighted sum based on stoichiometry or by using a neural network that can process variable-length sequences.

Diagram 1: atom2vec Featurization Workflow

Protocol 2: Training the SynthNN Model with PU Learning

Objective: To train a deep neural network to classify materials as synthesizable or unsynthesizable using a Positive-Unlabeled (PU) learning approach.

Background: A fundamental challenge in synthesizability prediction is the lack of verified negative examples (materials known to be unsynthesizable). The scientific literature primarily reports successful syntheses. PU learning addresses this by treating all materials not in the ICSD as unlabeled rather than definitively negative, and probabilistically reweights them during training based on the likelihood that they might be synthesizable [1] [22].

Materials:

• Featurized dataset from Protocol 1.
• Deep learning framework (e.g., TensorFlow/Keras or PyTorch).

Procedure:

• Construct Training Set:
  • Positive (P) Data: All featurized compositions from the ICSD.
  • Unlabeled (U) Data: Generate a large set of hypothetical chemical formulas (e.g., through combinatorial enumeration or perturbations of known formulas). These are the "artificially-generated unsynthesized materials" [1].
• Model Architecture: a. Design a deep neural network with an input layer that matches the dimensionality of the featurized composition vector. b. Include several fully connected (dense) hidden layers with non-linear activation functions (e.g., ReLU). c. The final output layer should be a single node with a sigmoid activation function to produce a synthesizability probability between 0 and 1.
• PU Learning Implementation: a. Implement a loss function that accounts for the PU learning scenario. A common approach is to use a weighted binary cross-entropy loss. b. Assign a weight of 1.0 to the positive examples (from ICSD). c. Assign a smaller, probabilistic weight (e.g., < 0.5) to the unlabeled examples, reflecting the prior belief that some of them may be synthesizable.
• Model Training: a. Split the P and U data into training and validation sets. b. Train the SynthNN model by minimizing the PU loss function. c. Use the validation set to monitor for overfitting and to tune hyperparameters (e.g., learning rate, network depth, PU weight).

Diagram 2: SynthNN PU Learning Architecture

Protocol 3: Benchmarking Against DFT and Human Experts

Objective: To quantitatively evaluate the performance of the trained SynthNN model against DFT-based formation energy calculations and the assessments of human material scientists.

Background: To validate its utility, SynthNN was subjected to a head-to-head material discovery challenge. The benchmark tests the model's precision in identifying plausible synthesizable materials from a large pool of candidates, comparing its efficiency and accuracy against established methods [1].

Materials:

• Trained SynthNN model from Protocol 2.
• A held-out test set of known and hypothetical materials.
• DFT software (e.g., VASP, Quantum ESPRESSO) for formation energy calculations [53] [54].
• A panel of expert solid-state chemists.

Procedure:

• Create Discovery Benchmark: a. Curate a diverse set of candidate chemical formulas, including some known synthesizable materials (from a reserved part of ICSD) and many hypothetical ones.
• Run SynthNN Screening: a. Use the trained SynthNN model to predict the synthesizability probability for every candidate in the benchmark set. b. Rank the candidates by their predicted probability and calculate the precision at various recall levels.
• Run DFT Screening: a. For each candidate, perform DFT calculations to obtain its formation energy with respect to its decomposition products [53] [55]. b. Use a formation energy threshold (e.g., energy above hull < 0 eV) to classify candidates as "stable" or "unstable." c. Calculate the precision of this DFT-based method on the benchmark set.
• Conduct Human Expert Assessment: a. Provide the same list of candidates to a panel of expert material scientists. b. Ask each expert to identify which materials they believe are synthesizable. c. Measure the precision and the time taken by each expert to complete the task.
• Analysis: a. Compare the precision of SynthNN, DFT, and the human experts. b. Document the massive disparity in computational time (seconds for SynthNN vs. hours/days for DFT) and human effort (seconds for SynthNN vs. days/weeks for experts).

Integration into a Discovery Pipeline

The true power of SynthNN is realized when it is integrated into a high-throughput computational screening workflow. As illustrated in the diagram below, SynthNN acts as a critical filter, ensuring that only the most synthetically promising candidates proceed to resource-intensive structure prediction and property calculation stages [1] [52]. This synthesizability constraint dramatically increases the hit rate and practical utility of virtual materials discovery campaigns.

Diagram 3: Synthesizability-Guided Discovery Pipeline

The application of machine learning (ML) in materials science and drug discovery relies heavily on effective numerical representations of atoms and molecules. Distributed representation methods, which encode chemical elements as dense vectors in a continuous space, have emerged as a powerful alternative to traditional hand-crafted features. This analysis provides a comparative examination of three prominent atom representation methods—Atom2Vec, Mat2Vec, and SkipAtom—with a specific focus on their application in predicting chemical synthesizability, a critical challenge in materials discovery and pharmaceutical development.

The core hypothesis underlying these methods is that atoms, like words in natural language, derive their "meaning" from their context. As articulated in the SkipAtom research, the approach formalizes the idea that "an atom shall be known by 'the company it keeps,'" aiming to learn chemical semantics from the chemo-structural contexts in which atoms appear [4] [50]. For researchers investigating synthesizability, these representations offer a data-driven pathway to identify promising compounds without relying exclusively on expensive density functional theory (DFT) calculations or human intuition.

Foundational Principles

Table 1: Core Methodological Approaches of Atom Representation Techniques

Method	Underlying Data Source	Core Algorithm	Dimensionality	Training Context
Atom2Vec	Materials database compositions	Co-occurrence matrix factorization (SVD)	Limited to number of atoms	Chemical environments in known compounds
Mat2Vec	Scientific abstracts (text corpus)	Word2Vec (Skip-gram/CBOW)	Typically 200 dimensions	Linguistic context in materials literature
SkipAtom	Crystal structure databases (e.g., Materials Project)	Skip-gram model with structural graphs	Typically 30-200 dimensions	Atomic connectivity in crystal structures

Atom2Vec Methodology

Atom2Vec employs an unsupervised approach to learn atomic representations by analyzing patterns in materials databases. The methodology involves generating a co-occurrence count matrix of atoms and their chemical environments from existing materials databases, followed by applying singular value decomposition (SVD) to this matrix to derive distributed atom vectors [4]. The resulting representations capture chemical similarity, with atoms that frequently appear in similar structural environments positioned closer together in the vector space. This method established that machines can autonomously learn fundamental properties of atoms from extensive compound databases, representing them as high-dimensional vectors that cluster elements in chemically meaningful ways [56].

Mat2Vec Methodology

Mat2Vec adapts the Word2Vec algorithm from natural language processing to the materials science domain. Instead of using crystal structures, it processes a textual corpus derived from millions of scientific abstracts related to materials science research [4]. The model learns to predict the context words surrounding target words, with the resulting embedding matrix capturing semantic relationships between materials science concepts, including elements. This approach benefits from the rich contextual information present in scientific language, where elements are discussed in relation to properties, applications, and synthesis conditions. The resulting vectors have been utilized in various materials informatics applications, achieving competitive performance in property prediction tasks [3].

SkipAtom Methodology

SkipAtom extends the NLP analogy more directly by using atomic connectivity in crystal structures as the training context. The approach represents crystal structures as graphs, where atoms are nodes and bonds are edges, then applies a Skip-gram-like objective to predict neighboring atoms given a target atom [4] [57]. Formally, the method maximizes the average log probability across all materials in a database:

Where M represents the set of materials, A_m the atoms in material m, and N(a) the neighbors of atom a [4]. The graph representation is typically derived using Voronoi decomposition, which identifies nearest neighbors using solid angle weights to determine coordination environment probabilities [4]. The resulting vectors capture nuanced chemo-structural relationships that reflect both chemical similarity and structural preferences.

Quantitative Performance Comparison

Benchmarking on Materials Property Prediction

Table 2: Performance Comparison on Material Property Prediction Tasks

Method	Formation Energy (MAE eV/atom)	Band Gap (MAE eV)	Synthesizability Prediction (Precision)	Remarks
Atom2Vec	Variable across tasks	Variable across tasks	Used in SynthNN (7× better than DFT) [1]	Performance depends on specific application
Mat2Vec	Competitive with structure-based benchmarks [4]	Competitive with structure-based benchmarks [4]	Not explicitly reported	Best performance in 4/8 benchmark tasks [4]
SkipAtom	0.08-0.12 (Elpasolite) [58]	Comparable to benchmarks [4]	Not explicitly reported	Best performance in 2/8 benchmark tasks [4]
One-hot Vectors	Higher errors compared to distributed representations [4]	Higher errors compared to distributed representations [4]	Not applicable	Baseline method

Independent evaluations demonstrate that Mat2Vec and SkipAtom representations generally outperform Atom2Vec on various property prediction tasks [50]. In comprehensive benchmarking across multiple datasets from the Matbench test suite, SkipAtom performed comparably to Mat2Vec and superior to Atom2Vec, with each method excelling in different domains [4] [50]. The performance advantage of distributed representations over traditional one-hot encodings is particularly pronounced for smaller datasets, where pre-trained embeddings provide valuable prior knowledge [50].

Application to Synthesizability Prediction

The prediction of synthesizability represents a particularly challenging application for representation methods. The SynthNN model exemplifies how these representations can be leveraged for synthesizability classification, achieving a 7× higher precision in identifying synthesizable materials compared to using DFT-calculated formation energies alone [1]. In a remarkable demonstration, SynthNN outperformed 20 expert materials scientists, achieving 1.5× higher precision and completing the task five orders of magnitude faster than the best human expert [1].

Without explicit programming of chemical rules, models using these representations automatically learn fundamental chemical principles including charge-balancing, chemical family relationships, and ionicity, utilizing these principles to generate synthesizability predictions [1]. This emergent capability highlights the power of data-driven representations to capture complex chemical relationships that are challenging to formalize explicitly.

Diagram 1: Workflow for Atom Representation Learning and Synthesizability Prediction. The diagram illustrates how different data sources feed into the three representation methods, which then support synthesizability prediction models like SynthNN.

Detailed Experimental Protocols

Protocol 1: Training SkipAtom Embeddings

Purpose: To generate distributed representations of atoms from crystal structure data.

Materials and Data Sources:

• Primary Data: Crystal structures from the Materials Project database (126,335 structures used in reference implementation) [57]
• Software: SkipAtom Python package (requires installation with pip install skipatom[training])
• Hardware: Standard computational workstation with sufficient RAM for processing graph data

Procedure:

• Data Acquisition: Download crystal structure data from Materials Project API or use provided dataset (mp_2020_10_09.pkl.gz)
• Graph Construction: Convert crystal structures to graphs using Voronoi decomposition with solid angle weights for neighbor identification [4]
• Pair Generation: Generate connected atom pairs from the structural graphs (yielding ~15 million pairs from Materials Project dataset) [50] [57]
• Model Training: Train SkipAtom embeddings using the skip-gram objective with negative sampling
• Induction Step: Apply optional induction to adjust vectors of under-represented atoms using chemical similarity [57]

Key Parameters:

• Embedding dimensionality: 30-200 dimensions
• Context window size: Immediate neighbors in crystal graph
• Training epochs: Until convergence (monitor loss)
• Negative samples: Typically 5-15

Protocol 2: Evaluating Representations on Synthesizability Prediction

Purpose: To assess the performance of different atom representations on synthesizability classification.

Materials and Data Sources:

• Positive Examples: Synthesized materials from Inorganic Crystal Structure Database (ICSD) [1]
• Negative Examples: Artificially generated unsynthesized materials (treated as unlabeled data in PU learning framework)
• Representations: Pre-trained Atom2Vec, Mat2Vec, and SkipAtom embeddings
• Model Architecture: Deep neural network with atom embedding matrix optimized alongside other parameters [1]

Procedure:

• Data Preparation:
  • Extract chemical formulas from ICSD for positive examples
  • Generate negative examples through combinatorial composition generation
  • Apply positive-unlabeled (PU) learning framework to account for incomplete labeling [1]

• Model Implementation:

  • Implement SynthNN-like architecture with embedding layer
  • Use atom2vec framework where chemical formulas are represented by learned atom embedding matrix [1]
  • Optimize embedding dimensionality as hyperparameter (typically 200 dimensions)

• Training:

  • Employ semi-supervised learning approach
  • Probabilistically reweight unlabeled examples according to likelihood of synthesizability [1]
  • Balance ratio of artificially generated formulas to synthesized formulas (Nsynth hyperparameter)

• Evaluation:

  • Compare against baseline methods (random guessing, charge-balancing)
  • Assess using precision-recall metrics and F1-score
  • Perform cross-validation to ensure robustness

Validation: Compare model predictions against expert human assessments and experimental synthesis outcomes where available.

The Scientist's Toolkit: Essential Research Reagents

Table 3: Essential Resources for Atom Representation Research

Resource Name	Type	Purpose/Function	Accessibility
Materials Project API	Database	Provides crystal structures and computed properties for training	Public access via API
ICSD (Inorganic Crystal Structure Database)	Database	Source of experimentally synthesized structures for positive examples	Licensed/Subscription
SkipAtom Python Package	Software	Implementation of SkipAtom for training and using embeddings	Open source (MIT license)
Mat2Vec Embeddings	Pre-trained Model	200-dimensional atom vectors trained on scientific abstracts	Publicly available
Atom2Vec Implementation	Software/Model	Code for generating atomic representations via SVD	Research code
OQMD (Open Quantum Materials Database)	Database	Benchmark dataset for formation energy prediction	Public access
Matbench	Benchmark Suite	Standardized tasks for evaluating materials ML models	Open source

Integration with Synthesizability Research

For researchers focusing specifically on synthesizability prediction, the representation method should be selected based on the specific requirements of the study. Atom2Vec has been directly validated in synthesizability prediction through the SynthNN model, which "leverages the entire space of synthesized inorganic chemical compositions" and identifies synthesizable materials with significantly higher precision than formation energy-based approaches [1].

The critical advantage of these representations for synthesizability research lies in their ability to learn implicit chemical rules without explicit programming. As noted in the SynthNN study, "without any prior chemical knowledge, our experiments indicate that SynthNN learns the chemical principles of charge-balancing, chemical family relationships and ionicity, and utilizes these principles to generate synthesizability predictions" [1].

Diagram 2: From Chemical Formula to Synthesizability Prediction. This workflow shows how atom representations are combined through pooling operations to create material-level representations for synthesizability classification.

The comparative analysis of Atom2Vec, Mat2Vec, and SkipAtom reveals a rapidly evolving landscape in atom representation learning for materials informatics. While each method demonstrates distinct strengths, SkipAtom and Mat2Vec generally outperform Atom2Vec on standard benchmark tasks, with the specific advantage depending on the target application [4] [50].

For synthesizability prediction specifically, Atom2Vec has shown remarkable efficacy when integrated into the SynthNN framework, outperforming both computational proxies like formation energy and human experts in identifying synthesizable compositions [1]. This success underscores the value of data-driven representations that capture complex chemical relationships beyond simple heuristics like charge balancing, which alone captures only a fraction of known synthesized compounds [1].

Future research directions include developing representations that account for oxidation states and coordination environments, integrating multi-modal information from both structural databases and scientific literature, and creating task-specific representation learning approaches optimized for synthesizability prediction. As these methods mature, they promise to significantly accelerate the discovery of novel materials by providing more reliable assessments of synthetic accessibility before experimental investment.

The acceleration of materials design through computational and data-driven paradigms has identified millions of candidate materials with promising properties. However, a significant bottleneck remains: many theoretically predicted crystal structures are not synthetically accessible, creating a critical gap between computational design and real-world application [59]. Traditional methods for assessing synthesizability, such as evaluating thermodynamic formation energies or kinetic stability via phonon spectra, often show poor correlation with actual synthesizability, as many metastable structures are synthesizable while numerous thermodynamically stable structures are not [59] [1].

The emergence of large language models (LLMs) offers a transformative approach to this challenge. This Application Note evaluates the Crystal Synthesis Large Language Models (CSLLM) framework, a novel application of specialized LLMs designed to accurately predict the synthesizability of arbitrary 3D crystal structures, their likely synthetic methods, and suitable precursors [59] [60]. The content is framed within the broader context of representational learning for materials, particularly building upon the concept of atom2vec and other distributed representations of atoms and materials for machine learning [4].

The CSLLM Framework: Architecture and Components

The CSLLM framework addresses the challenge of crystal structure synthesizability through three specialized LLMs, each fine-tuned for a specific sub-task [59]:

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

This tripartite architecture moves beyond traditional binary classification to provide a comprehensive synthesis planning tool. The framework processes crystal structures represented as text strings, enabling the application of advanced natural language processing techniques to materials science challenges [59].

Workflow Visualization

The following diagram illustrates the complete CSLLM framework workflow, from data preparation to synthesizability and precursor predictions:

Quantitative Performance Assessment

Comparative Performance Metrics

Table 1: Performance comparison of CSLLM against traditional synthesizability assessment methods

Method	Accuracy (%)	Advantage over Thermodynamic (%)	Advantage over Kinetic (%)
CSLLM (Synthesizability LLM)	98.6	106.1	44.5
Traditional Thermodynamic (Energy above hull ≥0.1 eV/atom)	74.1	-	-
Traditional Kinetic (Lowest phonon frequency ≥ -0.1 THz)	82.2	-	-
Method LLM	91.0 (Classification Accuracy)	N/A	N/A
Precursor LLM	80.2 (Success Rate)	N/A	N/A

The Synthesizability LLM demonstrates remarkable accuracy, significantly outperforming traditional approaches based on thermodynamic and kinetic stability [59] [60]. This performance advantage is particularly notable given that these traditional methods have been the standard for synthesizability screening in computational materials science.

Dataset Composition and Model Generalization

Table 2: CSLLM training dataset composition and model performance characteristics

Parameter	Value	Additional Context
Synthesizable Structures	70,120	From ICSD, ≤40 atoms, ≤7 elements [59]
Non-synthesizable Structures	80,000	Lowest CLscore structures from 1.4M theoretical pool [59]
Total Training Data	150,120	Balanced comprehensive dataset [59]
Generalization Accuracy	97.9%	On complex structures with large unit cells [59]
Positive Example Validation	98.3%	Percentage of ICSD structures with CLscore >0.1 [59]

The framework's exceptional generalization capability is demonstrated by its 97.9% accuracy on experimental structures with complexity considerably exceeding that of the training data [59]. This suggests that the model learns fundamental principles of materials synthesis rather than merely memorizing training examples.

Experimental Protocols and Methodologies

Protocol 1: Dataset Construction for Synthesizability Prediction

Purpose: To construct a balanced, comprehensive dataset of synthesizable and non-synthesizable crystal structures for training the Synthesizability LLM.

Materials and Input Data:

• Inorganic Crystal Structure Database (ICSD) [59] [1]
• Theoretical structures from Materials Project, Computational Material Database, Open Quantum Materials Database, and JARVIS [59]

Procedure:

• Positive Example Selection:
  • Extract 70,120 crystal structures from ICSD
  • Apply filters: maximum 40 atoms per structure, maximum 7 different elements
  • Exclude disordered structures to focus on ordered crystal structures [59]

• Negative Example Selection:

  • Collect 1,401,562 theoretical crystal structures from multiple databases
  • Calculate CLscore for each structure using pre-trained PU learning model [59]
  • Select 80,000 structures with lowest CLscores (CLscore <0.1) as non-synthesizable examples [59]

• Dataset Validation:

  • Compute CLscores for positive examples to validate threshold
  • Confirm 98.3% of positive examples have CLscore >0.1 [59]
  • Visualize dataset coverage using t-SNE to ensure comprehensive representation of crystal systems and elements [59]

Protocol 2: Material String Representation Development

Purpose: To create an efficient text representation for crystal structures that enables effective fine-tuning of LLMs.

Background: Traditional crystal structure representations like CIF or POSCAR contain redundant information and are not optimized for LLM processing [59].

Procedure:

• Essential Information Extraction:
  • Extract space group symbol and symmetry information
  • Collect lattice parameters (a, b, c, α, β, γ)
  • Identify unique atomic sites with their Wyckoff positions [59]

• Material String Construction:

  • Develop compact text representation: "SP | a, b, c, α, β, γ | (AS1-WS1[WP1-x,y,z]; ...)" [59]
  • Where SP: space group symbol, a,b,c,α,β,γ: lattice parameters, AS: atomic symbol, WS: Wyckoff symbol, WP: Wyckoff position with coordinates [59]

• Representation Validation:

  • Ensure reversible encoding (ability to reconstruct essential crystal information)
  • Verify information completeness for synthesizability assessment
  • Test computational efficiency for LLM processing [59]

Protocol 3: LLM Fine-Tuning for Synthesis Prediction

Purpose: To adapt general-purpose LLMs for specialized crystal synthesis prediction tasks.

Procedure:

• Model Architecture Selection:
  • Utilize three separate LLM instances for specialized tasks [59]
  • Employ domain-focused fine-tuning to align linguistic features with material features [59]

• Fine-Tuning Process:

  • Use constructed dataset of 150,120 text-represented crystal structures
  • Apply efficient fine-tuning techniques to adapt attention mechanisms to material science context [59]
  • Implement strategies to reduce "hallucination" and improve reliability [59]

• Model Validation:

  • Evaluate Synthesizability LLM on separate testing dataset
  • Assess Method LLM classification accuracy for synthetic pathways
  • Validate Precursor LLM through success rate in identifying appropriate precursors for binary and ternary compounds [59]

Material String Representation

The following diagram illustrates the process of converting a crystal structure into the specialized "material string" representation used by CSLLM:

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential computational tools and databases for crystal synthesizability research

Tool/Database	Type	Primary Function	Relevance to CSLLM
Inorganic Crystal Structure Database (ICSD)	Database	Repository of experimentally synthesized crystal structures [59] [1]	Source of positive (synthesizable) training examples
Materials Project	Database	Repository of computationally predicted crystal structures [59]	Source of candidate structures for negative examples
CLscore Model	Computational Model	Positive-unlabeled learning model for synthesizability scoring [59]	Screening theoretical structures for non-synthesizable examples
Material String Representation	Data Format	Text-based crystal structure representation [59]	Enables LLM processing of crystal structures
Atom2Vec/SkipAtom	Representation Learning	Unsupervised learning of distributed atom representations [4]	Provides foundational atomic embeddings for materials informatics
Composition Analyzer Featurizer (CAF)	Featurization Tool	Generates numerical compositional features from chemical formulas [3]	Complementary approach for composition-based prediction

Application and Implementation

The CSLLM framework has been successfully applied to screen 105,321 theoretical structures, identifying 45,632 as synthesizable [59]. These synthesizable candidates were further analyzed using graph neural network models to predict 23 key properties, enabling efficient prioritization for experimental synthesis [59].

A user-friendly graphical interface has been developed to facilitate broad adoption, allowing researchers to upload crystal structure files and automatically receive synthesizability predictions and precursor recommendations [59] [60]. This implementation significantly lowers the barrier for integrating CSLLM into computational materials discovery workflows.

The CSLLM framework represents a significant advancement in applying large language models to the critical challenge of crystal structure synthesizability prediction. By achieving 98.6% accuracy and demonstrating exceptional generalization capabilities, it substantially outperforms traditional thermodynamic and kinetic stability assessments. The framework's comprehensive approach—encompassing synthesizability prediction, method classification, and precursor identification—bridges a crucial gap between theoretical materials design and experimental synthesis. When contextualized within the broader field of atom2vec representation research, CSLLM exemplifies the powerful synergy between distributed material representations and advanced natural language processing, paving the way for accelerated discovery of novel functional materials.

The discovery of novel inorganic crystalline materials begins with identifying synthesizable chemical compositions—materials that are synthetically accessible through current capabilities, regardless of whether they have been synthesized yet [1]. Predicting synthesizability presents a classic binary classification challenge in machine learning, where models must distinguish between synthesizable and unsynthesizable materials. Unlike organic molecules that can often be synthesized through established reaction sequences, targeted synthesis of crystalline inorganic materials is complicated by the lack of well-understood reaction mechanisms [1]. This complexity necessitates sophisticated evaluation metrics that can reliably assess model performance under conditions of extreme class imbalance and uncertain negative examples.

Within the context of atom2vec representation for chemical formulas, performance metrics take on additional significance. The atom2vec approach represents each chemical formula through a learned atom embedding matrix optimized alongside other neural network parameters, learning an optimal representation directly from the distribution of previously synthesized materials without pre-defined chemical assumptions [1]. This representation reformulates materials discovery as a synthesizability classification task, requiring metrics that can accurately reflect model performance for integration into computational materials screening workflows. The choice of appropriate metrics—particularly precision, recall, and AUROC (Area Under the Receiver Operating Characteristic Curve)—directly impacts the reliability of synthesizability predictions and ultimately determines the success of autonomous materials discovery efforts.

Fundamental Metric Definitions and Calculations

Core Metrics from the Confusion Matrix

All binary classification metrics for synthesizability prediction derive from the confusion matrix, which specifies the relationship between ground truth labels and model predictions at a particular classification threshold [61]. The confusion matrix consists of four key components: True Positives (TPs), True Negatives (TNs), False Positives (FPs), and False Negatives (FNs) [62]. From these components, the fundamental metrics of precision and recall are calculated.

Precision (also called Positive Predictive Value) measures the accuracy of positive predictions, calculated as TP/(TP + FP) [61] [62]. In synthesizability contexts, precision answers the question: "When the model predicts a material as synthesizable, how often is it correct?" High precision indicates that the model minimizes false positives, which is crucial when experimental validation resources are limited.

Recall (also called Sensitivity or True Positive Rate) measures the model's ability to identify all actual positive instances, calculated as TP/(TP + FN) [61] [62]. For synthesizability prediction, recall answers: "Of all truly synthesizable materials, what proportion does the model successfully identify?" High recall ensures that potentially valuable materials aren't overlooked in the discovery process.

Threshold-Independent Metrics: AUROC and PR-AUC

The Receiver Operating Characteristic (ROC) curve visualizes the trade-off between the True Positive Rate (recall) and False Positive Rate (FPR) across all possible classification thresholds [63] [64]. The False Positive Rate is calculated as FP/(FP + TN) [61]. The Area Under the ROC Curve (AUROC) provides a single value summarizing the model's overall classification performance, with 1.0 representing a perfect classifier and 0.5 representing random guessing [62] [64].

The Precision-Recall (PR) curve visualizes the direct trade-off between precision (y-axis) and recall (x-axis) at various threshold settings [63] [65]. The Area Under the PR Curve (PR-AUC) summarizes this relationship, with higher values indicating better performance [65]. Unlike AUROC, the baseline PR-AUC for a random classifier equals the class imbalance ratio (proportion of positive examples) in the dataset [61].

Table 1: Fundamental Binary Classification Metrics for Synthesizability Prediction

Metric	Formula	Interpretation in Synthesizability Context
True Positive (TP)	-	Correctly identified synthesizable materials
False Positive (FP)	-	Unsynthesizable materials incorrectly labeled as synthesizable
True Negative (TN)	-	Correctly identified unsynthesizable materials
False Negative (FN)	-	Synthesizable materials incorrectly labeled as unsynthesizable
Precision	TP / (TP + FP)	Proportion of predicted-synthesizable materials that are actually synthesizable
Recall (Sensitivity)	TP / (TP + FN)	Proportion of actual synthesizable materials that are correctly identified
False Positive Rate	FP / (FP + TN)	Proportion of unsynthesizable materials incorrectly flagged as synthesizable
AUROC	Area under ROC curve	Overall measure of classification performance across all thresholds
PR-AUC	Area under PR curve	Model performance focused on the positive (synthesizable) class

Metric Selection for Imbalanced Synthesizability Data

The Class Imbalance Challenge in Materials Science

Synthesizability prediction inherently involves extreme class imbalance, with synthesizable materials representing only a tiny fraction of possible chemical compositions [1]. Traditional metrics like accuracy become misleading in such contexts, as a model that always predicts "unsynthesizable" could achieve high accuracy while being useless for materials discovery [62] [65]. This imbalance necessitates careful metric selection aligned with research objectives.

Research indicates that AUROC is robust to class imbalance when the score distribution remains unchanged, as it measures the model's ranking ability rather than absolute classification performance at a specific threshold [61]. The random baseline for AUROC is always 0.5, regardless of imbalance. Conversely, PR-AUC is highly sensitive to class imbalance, with its random baseline equal to the proportion of positive examples in the dataset [61]. This sensitivity makes PR-AUC particularly useful for evaluating performance on the minority class of interest.

When to Prefer PR-AUC Over AUROC

PR-AUC is preferable when the positive class (synthesizable materials) is the primary focus and the dataset is imbalanced [66] [65]. This aligns perfectly with synthesizability prediction, where researchers are more concerned with correctly identifying synthesizable materials than with correctly rejecting unsynthesizable ones. PR-AUC provides a more informative performance measure than AUROC in these contexts because it focuses specifically on the model's performance on the positive class [61] [65].

For the SynthNN model (a deep learning synthesizability model leveraging atom2vec representations), the precision-recall framework is particularly valuable due to the positive-unlabeled learning approach, where artificially generated unsynthesized materials may include some synthesizable compounds that simply haven't been discovered yet [1]. This creates inherent uncertainty in negative examples, making PR-AUC a more reliable metric than metrics assuming definitively labeled negatives.

Table 2: Metric Selection Guidelines for Synthesizability Prediction

Scenario	Recommended Primary Metric	Rationale
Initial model screening	AUROC	Provides overall performance assessment independent of class imbalance
Focus on synthesizable materials	PR-AUC	Emphasizes performance on the positive (minority) class
High-cost experimental validation	Precision	Minimizes false positives to reduce wasted resources
Comprehensive materials discovery	Recall	Maximizes identification of all synthesizable materials
Comparison across datasets	AUROC	Invariant to different class imbalances across datasets
Deploying operational workflow	F1 Score	Balances precision and recall for practical implementation

Experimental Protocols for Metric Evaluation

Protocol 1: Synthesizability Classification with Atom2Vec

Purpose: To evaluate synthesizability classification performance using atom2vec representations and standard metrics.

Materials and Data Sources:

• Inorganic Crystal Structure Database (ICSD): Source of positive examples (synthesized materials) [1]
• Artificially generated compositions: Source of negative examples (potentially unsynthesized materials) [1]
• Atom2Vec implementation: For learning element representations from chemical formulas [1] [3]
• SynthNN architecture: Deep learning model for synthesizability classification [1]

Procedure:

• Data Preparation: Extract known inorganic crystalline materials from ICSD as positive examples [1]
• Negative Example Generation: Create artificially generated chemical formulas as negative examples, acknowledging that some may be synthesizable but undiscovered [1]
• Representation Learning: Implement atom2vec to learn element embeddings from the entire space of synthesized inorganic chemical compositions [1]
• Model Training: Train SynthNN using a semi-supervised positive-unlabeled learning approach, treating unsynthesized materials as unlabeled data and probabilistically reweighting them according to likelihood of synthesizability [1]
• Metric Calculation:
  • Generate predicted probabilities for the test set
  • Compute precision-recall curves using sklearn.metrics.precision_recall_curve [65]
  • Calculate ROC curves using sklearn.metrics.roc_curve [64]
  • Compute AUROC and PR-AUC values using sklearn.metrics.auc [65]
• Performance Benchmarking: Compare against baseline methods including random guessing and charge-balancing approaches [1]

Synthesizability Classification Workflow

Protocol 2: Threshold Optimization for Practical Deployment

Purpose: To determine optimal classification thresholds for operational synthesizability prediction workflows.

Materials:

• Trained synthesizability classification model (e.g., SynthNN)
• Validation set with known synthesizability labels
• Computational resources for threshold sweeping

Procedure:

• Probability Prediction: Obtain predicted probabilities for the positive class (synthesizable) on the validation set
• Threshold Sweeping: Iterate over classification thresholds from 0 to 1.0 in increments of 0.01
• Metric Calculation at Each Threshold:
  • Convert probabilities to binary predictions using current threshold
  • Compute confusion matrix (TP, FP, TN, FN)
  • Calculate precision, recall, F1-score (harmonic mean of precision and recall) [63] [62]
• F1-Score Maximization: Identify the threshold that maximizes F1-score for balanced precision and recall [63]
• Precision-Focused Tuning: For resource-constrained scenarios, identify the threshold that achieves target precision (e.g., 90%)
• Recall-Focused Tuning: For comprehensive discovery, identify the threshold that achieves target recall (e.g., 90%)
• Validation: Apply selected thresholds to held-out test set and compare performance

The Scientist's Toolkit: Research Reagents and Computational Materials

Table 3: Essential Research Tools for Synthesizability Prediction Research

Tool/Resource	Type	Function in Synthesizability Research
Inorganic Crystal Structure Database (ICSD)	Data Resource	Primary source of synthesized materials for training and benchmarking [1]
Atom2Vec	Algorithm	Learns optimal element representations from chemical formula data [1] [3]
SynthNN	Model Architecture	Deep learning classifier for synthesizability prediction [1]
Precrec Library (R)	Software	Fast and accurate precision-recall calculations for large datasets [66]
scikit-learn Metrics	Software	Python implementation of standard classification metrics [63] [62] [65]
Materials Project Database	Data Resource	Source of computed material properties for feature engineering [3]
Matminer	Software Toolkit	Featurization toolkit for materials science data [3]
BLMM Crystal Transformer	Model Architecture	Blank-filling language model for generative materials design [7]

Interpreting Metrics in Research Context

Benchmarking Against Baselines and Human Experts

Meaningful interpretation of synthesizability metrics requires comparison against appropriate baselines. The random guessing baseline for AUROC is 0.5, while for PR-AUC it equals the proportion of positive examples in the dataset [61]. Charge-balancing—a commonly used heuristic in materials discovery—provides a chemically informed baseline, though it identifies only 37% of known synthesized materials as charge-balanced [1].

Remarkably, the SynthNN model with atom2vec representation has demonstrated superior performance compared to human experts, achieving 1.5× higher precision in material discovery tasks while completing the task five orders of magnitude faster than the best human expert [1]. This highlights the practical value of robust metric evaluation in synthesizability prediction.

Metric Trade-offs and Research Objectives

Different research objectives necessitate emphasis on different metrics. When prioritizing efficient use of experimental resources, precision should be emphasized to minimize false positives. When conducting comprehensive exploration of chemical space, recall becomes more important to minimize false negatives. The F1-score provides a balanced measure when both precision and recall are valued equally [63] [62].

Metric Selection Based on Research Goals

Advanced Considerations in Metric Implementation

Positive-Unlabeled Learning in Synthesizability Prediction

The positive-unlabeled (PU) learning framework is particularly relevant for synthesizability prediction, as definitive negative examples (provably unsynthesizable materials) are rarely available [1]. Instead, datasets typically contain confirmed positive examples (synthesized materials) and unlabeled examples that may include both positive and negative instances. This framing changes metric interpretation, as "false positives" may include synthesizable materials that simply haven't been synthesized yet.

In PU learning contexts, precision estimates may be artificially lowered, as some examples classified as false positives might actually be undiscovered positive examples [1]. Thus, relative performance comparisons between models may be more reliable than absolute metric values. The F1-score is often used as the primary evaluation metric for PU learning algorithms, as it balances the trade-off between precision and recall despite the uncertain labeling [1].

Computational Efficiency in Large-Scale Screening

Computational efficiency in metric calculation becomes crucial when screening billions of candidate materials [1]. For large-scale applications, specialized libraries like Precrec provide fast and accurate precision-recall calculations, using optimized algorithms and C++ implementations to handle massive datasets efficiently [66]. These implementations correctly handle non-linear interpolation between points on precision-recall curves, which is essential for accurate AUC calculations [66].

When working with extremely large chemical spaces, approximate metric calculations may be necessary. Sampling techniques can provide statistically valid estimates of performance metrics while reducing computational requirements. In such cases, reporting confidence intervals alongside point estimates provides a more complete picture of model performance.

The acceleration of materials discovery hinges on the ability to reliably predict whether a proposed inorganic crystalline material is synthesizable. Traditional computational screenings, which often rely on density-functional theory (DFT)-calculated formation energies as a proxy for stability, capture only approximately 50% of synthesized materials [1]. The development of machine learning models that leverage advanced atomic representations, such as atom2vec, offers a transformative approach by learning the underlying chemistry of synthesizability directly from extensive databases of known materials [1] [4]. This application note details the experimental protocols and presents real-world validation data demonstrating the successful application of these models in predicting the synthesizability of both known and novel materials.

Key Research Reagent Solutions

The following table catalogues the essential computational tools and data resources that form the foundation of synthesizability prediction research.

Table 1: Essential Research Reagents and Resources for Synthesizability Prediction

Resource Name	Type	Primary Function	Key Application
Inorganic Crystal Structure Database (ICSD) [1] [22]	Materials Database	Provides a curated collection of experimentally synthesized and characterized inorganic crystal structures.	Serves as the primary source of positive (synthesizable) examples for model training and validation.
Atom2Vec [1] [4]	Atomic Representation	Generates distributed vector representations (embeddings) of atoms by analyzing co-occurrence in known materials.	Learns chemical principles like charge-balancing and ionicity from data, enabling composition-based synthesizability classification.
SynthNN [1]	Deep Learning Model	A deep learning synthesizability model that uses atom embeddings to classify materials as synthesizable or not.	Directly predicts synthesizability from chemical composition alone, outperforming traditional stability metrics.
Positive-Unlabeled (PU) Learning [1] [22]	Machine Learning Framework	A semi-supervised learning paradigm designed for scenarios where only positive (synthesizable) examples are definitively known.	Handles the lack of confirmed negative data by treating unsynthesized materials as unlabeled.
Crystal Synthesis LLMs (CSLLM) [22]	Large Language Model Framework	A framework of fine-tuned LLMs that predict synthesizability, synthetic methods, and precursors from crystal structure text representations.	Achieves state-of-the-art accuracy in synthesizability prediction and provides actionable synthesis guidance.

Quantitative Performance Validation

Performance Against Traditional Methods

Extensive benchmarking validates that models like SynthNN significantly outperform traditional computational and human-based approaches for identifying synthesizable materials.

Table 2: Comparative Performance of Synthesizability Prediction Methods

Method	Key Principle	Reported Precision	Key Advantage/Limitation
SynthNN (Atom2Vec-based) [1]	Data-driven classification using learned atomic embeddings.	7x higher than DFT formation energy	Learns complex chemical relationships; does not require prior structural knowledge.
DFT Formation Energy [1] [22]	Thermodynamic stability relative to decomposition products.	~50% of synthesized materials	Fails to account for kinetic stabilization and non-thermodynamic factors.
Charge-Balancing [1]	Net neutral ionic charge based on common oxidation states.	Only 37% of known compounds	Inflexible; performs poorly for metallic, covalent, or complex ionic materials.
Human Expert [1]	Domain knowledge and specialized experience.	1.5x lower than SynthNN	Inherently limited in scope and speed compared to automated ML screening.
CSLLM Framework [22]	LLM fine-tuned on comprehensive dataset of material structures.	98.6% Accuracy	Requires crystal structure as input; suggests synthesis methods and precursors.

Case Study: SynthNN Validation

The SynthNN model, which leverages the atom2vec representation, was rigorously validated in a head-to-head test against human experts. The model achieved 1.5× higher precision in identifying synthesizable materials and completed the discovery task five orders of magnitude faster than the best-performing human expert [1]. Remarkably, despite being provided no explicit chemical rules, analysis of the model's predictions indicated that it had independently learned fundamental chemical principles such as charge-balancing, chemical family relationships, and ionicity [1].

Experimental Protocols

Protocol 1: Building a Synthesizability Dataset from ICSD

This protocol describes the creation of a dataset for training and testing synthesizability prediction models.

Purpose To compile a robust dataset of synthesizable and non-synthesizable inorganic crystalline materials from the Inorganic Crystal Structure Database (ICSD) and complementary theoretical databases.

Materials and Reagents

• ICSD Access: License to the Inorganic Crystal Structure Database [1] [22].
• Theoretical Databases: Access to materials repositories such as the Materials Project (MP), Open Quantum Materials Database (OQMD), or JARVIS [22].
• Computing Environment: Standard computer with Python and necessary data processing libraries (e.g., Pandas, NumPy).

Procedure

• Extract Positive Examples:
  • Query the ICSD for all reported, structurally characterized, inorganic crystalline materials.
  • Apply filters as needed for your study, for example, excluding disordered structures and limiting to compositions with ≤ 40 atoms and ≤ 7 distinct elements [22].
  • This curated list of ~70,000 materials constitutes the positive (synthesizable) class.

• Generate Negative Examples:

  • Compile a large set of theoretical crystal structures from databases like MP, OQMD, and JARVIS. A pool of over 1.4 million structures is typical [22].
  • To minimize false negatives, employ a pre-trained Positive-Unlabeled (PU) learning model to calculate a synthesizability confidence score for each structure.
  • Select structures with the lowest confidence scores as the negative (non-synthesizable) class. A common threshold is a score < 0.1, yielding ~80,000 negative examples [22].

• Data Validation:

  • Verify that a high percentage (>98%) of your ICSD-derived positive examples receive high confidence scores from the PU model, confirming the validity of the negative set construction [22].
  • Use dimensionality reduction techniques like t-SNE to visualize the dataset and confirm it covers diverse crystal systems and chemical spaces [22].

Notes Definitively labeling a material as "unsynthesizable" is challenging. The PU learning framework probabilistically handles this uncertainty [1].

Protocol 2: Training an Atom2Vec-Based Prediction Model (SynthNN)

This protocol outlines the steps for training a deep learning model to predict synthesizability from chemical compositions.

Purpose To train a model that leverages learned atom embeddings to accurately classify the synthesizability of inorganic chemical formulas.

Materials and Reagents

• Software: Python with deep learning frameworks (e.g., TensorFlow, PyTorch).
• Data: The Synthesizability Dataset prepared in Protocol 1.
• Atom2Vec Implementation: Code for the atom2vec algorithm or a pre-trained embedding model [1] [4].

Procedure

• Featurization:
  • Represent each chemical formula in the dataset using the atom2vec framework.
  • The model will learn an embedding matrix for each element, which is optimized alongside the other network parameters during training [1].
  • This step transforms a chemical formula into a fixed-length, numerical vector representation.

• Model Architecture & Training:

  • Design a neural network architecture. The specific topology is a hyperparameter, but it typically consists of several fully connected (dense) layers following the embedding layer [1].
  • Train the model using a Positive-Unlabeled (PU) learning loss function. This approach treats the artificially generated negative examples as unlabeled data and reweights them according to their likelihood of being synthesizable [1].
  • The ratio of generated unsynthesized formulas to known synthesized formulas is a critical hyperparameter to tune.

• Model Validation:

  • Evaluate the trained model on a held-out test set from Protocol 1.
  • Benchmark its performance against established baselines, including random guessing, charge-balancing, and formation energy thresholds, using metrics like precision, recall, and F1-score [1] [22].

Notes Unlike traditional featurizers, atom2vec does not require pre-defined elemental properties. The model learns the relevant chemical representations directly from the data [1] [4].

Protocol 3: Validating Model Predictions Experimentally

The ultimate test of any predictive model is the successful synthesis of its novel predictions.

Purpose To experimentally verify the synthesizability of novel material compositions identified by the trained model.

Materials and Reagents

• Candidate List: A shortlist of novel material compositions predicted to be synthesizable with high confidence.
• Precursor Materials: High-purity solid-state precursors (e.g., oxides, carbonates, metals) as determined by precursor prediction models or phase diagram analysis.
• Laboratory Equipment: Standard solid-state synthesis equipment: mortar and pestle, high-temperature furnaces, controlled atmosphere boxes, and materials characterization tools (e.g., X-ray Diffractometer).

Procedure

• Precursor Selection:
  • For a given candidate composition, use a precursor prediction LLM to suggest suitable solid-state precursors [22].
  • Alternatively, consult phase diagrams to identify thermodynamically favorable reactant mixtures.

• Synthesis Execution:

  • Weigh out precursor powders in the appropriate stoichiometric ratios.
  • Mechanically mix and grind the powders thoroughly to ensure homogeneity.
  • Pelletize the powder mixture to increase intimacy of contact.
  • React the pellets in a furnace at an optimized temperature and duration, often under an inert atmosphere or in a sealed quartz tube to prevent oxidation [22].

• Characterization and Confirmation:

  • Analyze the resulting product using powder X-ray Diffraction (XRD).
  • Compare the experimental diffraction pattern with the pattern simulated from the predicted crystal structure.
  • A successful synthesis is confirmed by a strong match between the experimental and simulated XRD patterns, indicating the target material has been formed.

Notes Not all high-confidence predictions will synthesize on the first attempt. Iterative optimization of synthesis parameters (temperature, time, cooling rate) is often necessary.

Conclusion

The integration of Atom2Vec and advanced deep learning representations marks a paradigm shift in predicting chemical synthesizability. These models demonstrably outperform traditional metrics like charge-balancing and DFT-based formation energy, offering a data-driven path to navigate vast chemical spaces. By learning complex chemical principles directly from data, they achieve superior precision and speed, accelerating the discovery of functional materials and viable drug candidates. Future directions involve the fusion of compositional and structural data, the development of more interpretable models to guide synthetic chemists, and the expansion into more complex chemical spaces, such as biomolecules and polymers. Ultimately, these tools are poised to significantly shorten the development timeline in materials science and clinical research, transforming theoretical designs into tangible innovations.

References

• 1. Predicting the synthesizability of crystalline inorganic ... [https://www.nature.com/articles/s41524-023-01114-4]
• 2. Generate what you can make: achieving in-house ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11954305/]
• 3. Composition and structure analyzer/featurizer for explainable ... [https://pubs.rsc.org/en/content/articlehtml/2025/dd/d4dd00332b]
• 4. Distributed representations of atoms and materials for ... [https://www.nature.com/articles/s41524-022-00729-3]
• 5. Atom2Vec: learning atoms for materials discovery [https://arxiv.org/abs/1807.05617]
• 6. Element2Vec: Learning Chemical Element Representation ... [https://arxiv.org/html/2510.13916v2]
• 7. Crystal Composition Transformer: Self‐Learning Neural ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11423232/]
• 8. Science - C.Thang Nguyen [https://thangckt.github.io/blog/category/science/]
• 9. Inverse Design of Solid-State Materials via a Continuous ... [https://www.sciencedirect.com/science/article/pii/S2590238519301754]
• 10. Predicting Synthesizability using Machine Learning on ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9996807/]
• 11. Element2Vec: Build Chemical Element Representation ... [https://arxiv.org/html/2510.13916v1]
• 12. Novel inorganic crystal structures predicted using ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9197834/]
• 13. Machine learned synthesizability predictions aided by density functional theory | Communications Materials [https://www.nature.com/articles/s43246-022-00295-7]
• 14. PubChem atom environments | Journal of Cheminformatics [https://jcheminf.biomedcentral.com/articles/10.1186/s13321-015-0076-4]
• 15. Learning atoms for materials discovery - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC6048531/]
• 16. How AI is Transforming the Elemental Chart in 2025 [https://www.pageon.ai/blog/elemental-chart-ai]
• 17. A generative model for inorganic materials design - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC11922738/]
• 18. Predicting the synthesizability of crystalline from... inorganic materials [https://www.nature.com/articles/s41524-023-01114-4?error=cookies_not_supported&code=42e0df74-4f12-47b6-b24f-a44b58abbf6e]
• 19. GitHub - antoniuk1/ SynthNN [https://github.com/antoniuk1/SynthNN]
• 20. Predicting the Synthesizability of Crystalline from... Inorganic Materials [https://www.researchsquare.com/article/rs-2574875/v1]
• 21. Solid-state synthesizability predictions using positive ... [https://pubs.rsc.org/en/content/articlehtml/2025/dd/d5dd00065c]
• 22. Accurate prediction of synthesizability and precursors of 3D ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12264133/]
• 23. MatterGen: a generative model design for inorganic materials [https://arxiv.org/abs/2312.03687]
• 24. A recent survey on instance-dependent positive and unlabeled learning - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC11997483/]
• 25. PSPU: Enhanced Positive and Unlabeled Learning by Leveraging Pseudo Supervision [https://arxiv.org/html/2407.06698v1]
• 26. Developing a Semi-Supervised Approach Using a PU- ... [https://www.mdpi.com/1422-0067/25/15/8239]
• 27. Generate what you can make: achieving in-house ... [https://jcheminf.biomedcentral.com/articles/10.1186/s13321-024-00910-4]
• 28. Urban Institute Data Visualization style guide - GitHub Pages [http://urbaninstitute.github.io/graphics-styleguide/]
• 29. Structure motif–centric learning framework for inorganic ... [https://www.osti.gov/pages/biblio/1824888]
• 30. Group graph: a molecular graph representation with ... [https://jcheminf.biomedcentral.com/articles/10.1186/s13321-024-00933-x]
• 31. Enhancing property and activity prediction and ... [https://www.nature.com/articles/s42004-024-01155-w]
• 32. DeepSA: a deep-learning driven predictor of compound ... [https://jcheminf.biomedcentral.com/articles/10.1186/s13321-023-00771-3]
• 33. Molecular optimization using a conditional transformer for ... [https://www.nature.com/articles/s42004-025-01437-x]
• 34. An artificial intelligence accelerated virtual screening ... [https://www.nature.com/articles/s41467-024-52061-7]
• 35. A high throughput synthetic workflow for solid state synthesis ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10866347/]
• 36. SynCoTrain: A Dual Classifier PU-learning Framework ... [https://arxiv.org/html/2411.12011v1]
• 37. Estimating classification accuracy in positive-unlabeled ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6417800/]
• 38. Positive and Unlabeled Materials Machine Learning [https://towardsdatascience.com/positive-and-unlabeled-materials-machine-learning-8b216edea899/]
• 39. Solid-state synthesizability predictions using positive- ... [https://pubs.rsc.org/en/content/articlelanding/2025/dd/d5dd00065c]
• 40. Machine learning and quantum theory for new medicines [https://www.bbva.ch/en/blog/innovacion/tecnologia/machine-learning-y-cuantica-para-crear-nuevos-medicamentos.html]
• 41. An atomic feature representation model for machine learning [https://www.sciencedirect.com/science/article/abs/pii/S0927025621005590]
• 42. MIT Researchers Propose a New Way To Create ... [https://scitechdaily.com/mit-researchers-propose-a-new-way-to-create-synthesizable-molecules/]
• 43. Interpreting Black-Box Models: A Review on Explainable ... [https://link.springer.com/article/10.1007/s12559-023-10179-8]
• 44. Making sense of AI: bias, trust and transparency in pharma ... [https://www.drugtargetreview.com/article/183358/making-sense-of-ai-bias-trust-and-transparency-in-pharma-rd/]
• 45. Peering inside the black box by learning the relevance of ... [https://www.nature.com/articles/s41467-025-65863-0]
• 46. and Balancing in Engineering Simulations Accuracy Cost [https://scisimple.com/en/articles/2025-06-13-balancing-accuracy-and-cost-in-engineering-simulations--akygyrz]
• 47. Improvement: Improving Cost and... - FasterCapital Accuracy Cost [https://fastercapital.com/content/Cost-Accuracy-Improvement--Improving-Cost-and-Accuracy-of-Deep-Learning-Models.html]
• 48. Advances in metastable phase catalysis with ... [https://www.sciencedirect.com/science/article/abs/pii/S1369702125003876]
• 49. Predicting synthesizability - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC6760004/]
• 50. Distributed Representations of Atoms - Materialis.AI - Blog [https://blog.materialis.ai/distributed-representations-of-atoms.html]
• 51. Predicting the synthesizability of crystalline inorganic ... [https://www.osti.gov/pages/biblio/1996869]
• 52. A Synthesizability-Guided Pipeline for Materials Discovery [https://arxiv.org/html/2511.01790v1]
• 53. density functional theory - Formation Energy Calculations - Matter Modeling Stack Exchange [https://mattermodeling.stackexchange.com/questions/12448/formation-energy-calculations]
• 54. Materials Database from All-electron Hybrid Functional DFT Calculations | Scientific Data [https://www.nature.com/articles/s41597-025-05867-z]
• 55. thermodynamics - Why formation energy calculation is essential in DFT studies? - Physics Stack Exchange [https://physics.stackexchange.com/questions/240175/why-formation-energy-calculation-is-essential-in-dft-studies]
• 56. [PDF] Distributed representations of atoms and materials ... [https://www.semanticscholar.org/paper/e8520c9cde21eed199183fb4d4b5b16e946fe9bd]
• 57. lantunes/skipatom: Distributed representations of atoms, ... [https://github.com/lantunes/skipatom]
• 58. Distributed Representations of Atoms and Materials for ... [https://www.bohrium.com/paper-details/distributed-representations-of-atoms-and-materials-for-machine-learning/867760024728895546-108483]
• 59. Accurate prediction of synthesizability and precursors of 3D ... [https://www.nature.com/articles/s41467-025-61778-y]
• 60. [2407.07016] Is Large Language Model All You Need to ... [https://arxiv.org/abs/2407.07016]
• 61. The receiver operating characteristic curve accurately ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11240176/]
• 62. Understanding F1 Score, Accuracy, ROC-AUC & PR- ... [https://www.deepchecks.com/f1-score-accuracy-roc-auc-and-pr-auc-metrics-for-models/]
• 63. F1 Score vs ROC AUC vs Accuracy vs PR AUC [https://neptune.ai/blog/f1-score-accuracy-roc-auc-pr-auc]
• 64. Understanding AUC — ROC and Precision-Recall Curves [https://medium.com/@data.science.enthusiast/understanding-auc-roc-and-precision-recall-curves-ae9180eaf4f7]
• 65. Ultimate Guide to PR-AUC: Calculations, Uses, and ... [https://coralogix.com/ai-blog/ultimate-guide-to-pr-auc-calculations-uses-and-limitations/]
• 66. Precrec: fast and accurate precision–recall and ROC curve ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5408773/]