Semi-Supervised Learning for Material Synthesizability: A New Paradigm for Accelerated Discovery

Abstract

Predicting material synthesizability remains a critical bottleneck in the discovery pipeline, compounded by the scarcity of labeled experimental data. This article explores how semi-supervised learning (SSL) leverages both limited labeled data and abundant unlabeled data to overcome this challenge. We cover the foundational principles of SSL in materials science, detail cutting-edge methodologies from Positive-Unlabeled learning to co-training frameworks, and address key challenges like class imbalance and data quality. A comparative analysis validates SSL's performance against supervised and self-supervised approaches, providing researchers and development professionals with a comprehensive guide to deploying these techniques for efficient and reliable synthesizability prediction.

The Synthesizability Challenge and the SSL Solution

The accelerating design of novel compounds through computational methods has created a fundamental bottleneck: the experimental realization of predicted materials. This challenge, known as the "synthesis gap," separates the abundance of computationally identified candidates from their practical creation in the laboratory [1]. While thermodynamic stability is a foundational concept in materials design, synthesizability encompasses a far broader set of factors that determine whether a material can actually be made. Defining synthesizability requires moving beyond simple energy calculations at 0 K to include kinetic pathways, chemical heuristics, and processing conditions [1]. This article explores this comprehensive definition of synthesizability and details how semi-supervised learning (SSL) is emerging as a powerful tool to quantify it and guide experimental efforts, thereby narrowing the divide between virtual design and real-world materials.

The Multi-Faceted Nature of Synthesizability

Synthesizability is the probability that a material with a given composition and crystal structure can be realized through a known or plausible experimental synthesis route. It is not a binary property but a continuum influenced by multiple interdependent factors.

Key Determinants of Synthesizability

• Thermodynamic Stability: This remains a primary filter. A material that is highly unstable relative to other phases in its chemical system is unlikely to be synthesizable. However, stability is necessary but not sufficient. The evaluation must also consider the thermodynamic landscape, including the Gibbs free energy at relevant reaction conditions, not just internal energy at 0 K [1].
• Kinetic Accessibility: Many useful materials are metastable. Their synthesis is possible if kinetic barriers prevent their decomposition to the global thermodynamic minimum. Synthesizability, therefore, depends on the existence of a kinetic pathway with manageable energy barriers.
• Chemical Intuition and Rules: Heuristics such as charge neutrality, electronegativity balance, and coordination chemistry provide crucial, rapid screens for likely stable compounds [1].
• Experimental Synthesis Routes: The specific sequence of experimental steps—such as grinding, heating, dissolving, or centrifuging—defines a synthesis methodology (e.g., solid-state or hydrothermal) and directly determines which materials can be accessed [2].

Semi-Supervised Learning as a Bridge Across the Synthesis Gap

The problem of predicting synthesizability is ideally suited for semi-supervised learning. In materials science, we have a small amount of data from expensive, carefully executed experiments (labeled data) and a vast corpus of unprocessed scientific literature and hypothetical compositions (unlabeled data). SSL leverages both to build powerful predictive models.

SSL Framework for Synthesizability Prediction

A common SSL framework, as applied in materials science, involves a two-stage process of topic modeling followed by supervised classification. The workflow for this approach is illustrated below.

This methodology allows researchers to convert unstructured natural language from millions of scientific articles into a machine-readable format. Latent Dirichlet Allocation (LDA) automatically identifies keywords associated with specific experimental steps, which are then used as features to train a robust classifier like a Random Forest, achieving high accuracy even with modest amounts of labeled data [2].

Positive-Unlabeled Learning for Stoichiometry Prediction

Another powerful SSL variant is Positive-Unlabeled (PU) Learning, which is directly applicable to predicting the synthesizability of material stoichiometries. In this setup, the model learns from a set of known synthesizable compositions (positives) and a large set of hypothetical compositions with unknown synthesizability (unlabeled). A notable application of this method achieved a true positive rate of 83.4% and an estimated precision of 83.6% on a test dataset [3]. This model's ability to treat arbitrary elemental combinations allows for the construction of continuous synthesizability phase maps, which can guide the experimental exploration of new compositional spaces and has led to the discovery of new phases, such as Cu₄FeV₃O₁₃ [3].

Quantitative Performance of SSL Models for Synthesis Prediction

The effectiveness of SSL approaches is demonstrated by concrete performance metrics across different applications, from text classification to direct synthesizability prediction.

Table 1: Performance Metrics of Semi-Supervised Learning Models in Materials Research

Application	SSL Method	Key Metric	Performance	Reference
Synthesis Procedure Classification	LDA + Random Forest	F1 Score	~90% (with >3000 training samples)	[2]
Synthesizability of Stoichiometry	Positive-Unlabeled Learning	True Positive Rate (Recall)	83.4%	[3]
Synthesizability of Stoichiometry	Positive-Unlabeled Learning	Estimated Precision	83.6%	[3]

These quantitative results underscore that SSL models can achieve high accuracy and reliability, providing a actionable tool for prioritizing candidate materials for experimental synthesis.

Detailed Experimental Protocol: Classifying Synthesis Procedures

This protocol details the method for using semi-supervised learning to classify materials synthesis procedures from written text, as established by Kim et al. [2].

Objectives and Preparation

• Primary Objective: To accurately classify paragraphs of scientific text into categories of inorganic materials synthesis methodologies (e.g., solid-state, hydrothermal, sol-gel).
• Materials and Software: A computing environment with Python and libraries including scikit-learn, gensim (for LDA), and nltk (for natural language processing). Access to a large corpus of materials science literature (e.g., from PubMed, other scientific databases) is required.

Step-by-Step Procedure

• Data Collection and Preprocessing:

  • Gather a large collection of scientific article abstracts and/or full texts (e.g., >2 million paragraphs) [2]. This constitutes the unlabeled dataset.
  • For the labeled dataset, manually annotate a few thousand paragraphs into the target synthesis categories (e.g., 1000 each for solid-state, hydrothermal, and sol-gel) and a "none" category. This requires domain expertise.

• Unsupervised Topic Modeling with LDA:

  • Preprocess the text from the unlabeled corpus: tokenize, remove stop words, and lemmatize.
  • Apply the Latent Dirichlet Allocation algorithm to the processed corpus. A model with 200 topics is a feasible starting point [2].
  • Interpret and label the resulting topics (e.g., "milling," "sintering," "autoclaving") based on their highest-probability keywords.

• Feature Engineering:

  • For each paragraph in the labeled training set, infer its topic distribution using the trained LDA model.
  • Construct "topic n-grams"—the sequences of LDA-derived topics in adjacent sentences within a paragraph. These n-grams serve as the input feature vectors for the classifier [2].

• Supervised Classification with Random Forest:

  • Split the labeled dataset into training and validation sets.
  • Train a Random Forest classifier using the topic n-gram features from the training set. Utilize hyperparameter tuning (e.g., number of trees) on the validation set to optimize performance.
  • Validate the final model's performance using standard metrics like F1 score, precision, and recall (see Table 1).

• Analysis and Interpretation:

  • Analyze the trained Random Forest model to identify the most important topic n-grams for classifying each synthesis type. This reveals the "machine intuition" of synthesis procedures, such as the pattern "[ball-]milling" → "sintering" for solid-state reactions [2].

The Scientist's Toolkit: Key Reagents and Methods

This table lists essential materials and computational tools frequently employed in the research and development of SSL models for synthesizability prediction.

Table 2: Essential Research Reagents and Tools for SSL-driven Materials Synthesis

Item Name	Function/Description	Relevance to SSL and Synthesis
Precursor Powders	High-purity metal oxides, carbonates, or other salts used as starting materials for solid-state reactions.	Essential for experimental validation of predicted synthesizable compositions (e.g., CuO, Fe₂O₃, V₂O₅ for discovering Cu₄FeV₃O₁₃) [3].
Ball Mill	Apparatus used for grinding and mixing precursor powders to achieve homogeneity and reduce particle size.	Represents a key experimental step ("milling") identified by LDA topic modeling in synthesis texts [2].
High-Temperature Furnace	Equipment for heating mixed precursors at high temperatures (sintering/calcination) to facilitate solid-state diffusion and reaction.	Represents the "sintering" topic in LDA models and is a critical step in many synthesis workflows [2].
Scikit-learn	A popular open-source Python library for machine learning.	Provides the implementation for the Random Forest classifier and other ML utilities used in the supervised learning stage [2].
Gensim	A robust open-source vector space modeling and topic modeling toolkit in Python.	Used to implement the Latent Dirichlet Allocation (LDA) algorithm for unsupervised topic discovery from text corpora [2].
Physical Property and Synthesis Databases	Databases such as the ICSD (Inorganic Crystal Structure Database) containing known structures and synthesis information.	Serve as critical sources of "positive" data points for training and validating SSL-based synthesizability models [3].

Closing the synthesis gap is a central challenge in modern materials science. A narrow focus on thermodynamic stability is insufficient; a holistic view of synthesizability that incorporates kinetic, chemical, and procedural factors is required. Semi-supervised learning stands out as a particularly effective computational framework for this problem, as it efficiently leverages the vast, untapped knowledge within the scientific literature combined with limited expert-labeled data. By implementing the protocols and models described, researchers can systematically prioritize the most promising candidate materials for synthesis, thereby accelerating the discovery and deployment of new functional materials.

Application Note: Leveraging Semi-Supervised Learning to Overcome Data Scarcity in Materials Science

The discovery and synthesis of new functional materials are fundamentally constrained by the data scarcity problem, particularly the systematic absence of data from failed experiments in scientific literature and databases. This positive publication bias creates severely imbalanced datasets where unsuccessful synthesis attempts are dramatically underrepresented [2] [4]. For materials discovery, this imbalance manifests as a critical knowledge gap that impedes the accurate prediction of synthesizability and stability. The high cost of failed experiments—both in terms of resources and time—compounds this issue, as each unsuccessful attempt generates valuable data that rarely enters the public domain [5]. Semi-supervised learning (SSL) has emerged as a powerful computational framework to address this challenge by leveraging both limited labeled data and abundant unlabeled data to build predictive models that can accurately identify synthesizable materials.

Quantitative Landscape of Data Scarcity in Materials Databases

Table 1: Data Imbalance in Major Materials Databases Affecting Synthesizability Prediction

Database	Total Entries	Stable/Synthesizable Materials	Unstable/Unsynthesizable Materials	Key Implication
Materials Project [4]	~138,613	~127,273 (91.8%)	~11,340 (8.2%)	Supervised models biased toward stable materials
Inorganic Crystal Structure Database (ICSD) [4]	~200,000	Majority class	Minimal representation	Lacks systematic failure data for training
Cambridge Structural Database (CSD) [5]	>100,000 transition metal complexes	Successfully synthesized structures	No failed synthesis records	Incomplete synthesis procedure landscape

Table 2: Performance Comparison of Materials Prediction Models

Model Type	Application	Performance Metric	Result	Limitations
Supervised CGCNN (baseline) [4]	Formation energy prediction	Accuracy	Baseline reference	Strong bias toward negative formation energy samples
TSDNN (Semi-supervised) [4]	Formation energy prediction	Accuracy	10.3% improvement over baseline	Requires careful pseudo-labeling
PU Learning (Semi-supervised) [3]	Synthesizability prediction	True Positive Rate	83.4%	Precision: 83.6%
Random Forest + LDA [2]	Synthesis procedure classification	F1 Score	~90%	Requires ~3000 training paragraphs

Semi-Supervised Learning Frameworks for Materials Synthesizability

Teacher-Student Dual Neural Network (TSDNN) for Formation Energy Prediction

The TSDNN framework addresses data scarcity through a unique dual-network architecture that effectively exploits large amounts of unlabeled data [4]. This approach specifically tackles the dataset bias where most samples in materials databases are stable, synthesizable materials with negative formation energies. The teacher model provides pseudo-labels for unlabeled data, which the student model then learns from, creating an iterative improvement cycle that significantly enhances prediction accuracy for out-of-distribution samples, including unstable hypothetical materials.

Key advantages:

• Achieves 10.3% absolute accuracy improvement for formation energy classification compared to supervised CGCNN baseline [4]
• Increases true positive rate for synthesizability prediction from 87.9% to 92.9% using only 1/49 model parameters [4]
• Successfully identified 512 novel stable cubic structures with negative formation energies from 1000 candidate samples when combined with CubicGAN generator [4]

Positive-Unlabeled (PU) Learning for Synthesizability Prediction

PU learning represents another SSL approach specifically designed for scenarios where only positive (successful) and unlabeled examples are available, perfectly matching the materials data landscape [3]. This method enables the prediction of synthesizability for any given elemental stoichiometry by learning the hidden features of synthesizable compositions from available data.

Experimental validation:

• Guided experimental exploration of quaternary oxide compositional space (CuO, Fe₂O₃, V₂O₅) [3]
• Resulted in discovery of new phase: Cu₄FeV₃O₁₃ [3]
• Demonstrates 83.4% true positive rate with estimated precision of 83.6% for test dataset [3]

Experimental Protocols

Protocol 1: Implementing Teacher-Student Dual Neural Network for Formation Energy Prediction

Data Preparation and Preprocessing

• Data Collection: Gather labeled and unlabeled materials data from sources including:

  • Materials Project API (formation energies, crystal structures) [4]
  • ICSD (synthesized crystal structures) [4]
  • Unlabeled hypothetical materials from generative models

• Feature Representation:

  • Convert crystal structures to graph representations using Crystal Graph Convolutional Neural Network (CGCNN) framework [4]
  • Node features: atom attributes (element type, formal charge, etc.)
  • Edge features: crystallographic bond characteristics

• Dataset Splitting:

  • Labeled training set: 8-10% of available data with known formation energies [4]
  • Unlabeled set: Remaining 90%+ of materials data
  • Standard train/validation/test splits (e.g., 80/10/10) for evaluation

Model Architecture and Training

• Training Procedure:

  • Phase 1: Pre-train teacher model on limited labeled data
  • Phase 2: Generate pseudo-labels for unlabeled data using teacher model
  • Phase 3: Train student model on combined labeled data and pseudo-labeled data
  • Phase 4: Apply consistency regularization to improve teacher model
  • Phase 5: Iterate steps 2-4 until convergence

• Hyperparameter Optimization:

  • Learning rate: 0.001-0.01 with decay schedule
  • Batch size: 32-128 depending on available memory
  • Number of CGCNN convolution layers: 3-5
  • Hidden layer dimensions: 64-128 neurons

Model Validation and Evaluation

• Performance Metrics:

  • Formation energy prediction: Mean Absolute Error (MAE)
  • Stability classification: Precision, Recall, F1-score, Accuracy
  • Synthesizability prediction: True Positive Rate, Precision

• Validation Techniques:

  • Cross-validation on labeled data
  • Ablation studies to assess contribution of unlabeled data
  • Comparison against supervised baselines (CGCNN, etc.)

Protocol 2: Natural Language Processing for Synthesis Procedure Classification

Text Processing and Feature Extraction

• Corpus Collection:

  • Gather 2,284,577 scientific articles from materials science literature [2]
  • Extract synthesis procedure paragraphs and experimental sections

• Unsupervised Topic Modeling:

  • Apply Latent Dirichlet Allocation (LDA) to identify experimental steps [2]
  • Cluster keywords into topics corresponding to specific synthesis steps
  • Identify 200 topics representing common synthesis operations [2]

• Feature Engineering:

  • Generate "topic n-grams" representing sequences of experimental steps [2]
  • Create document-topic distributions for each synthesis paragraph

Semi-Supervised Classification

• Model Training:

  • Annotate 1000 training paragraphs for each synthesis type (solid-state, hydrothermal, sol-gel) [2]
  • Include 3000 negative paragraphs without these synthesis procedures
  • Train Random Forest classifier on topic n-gram features
  • Utilize ensemble of 20 decision trees for optimal performance [2]

• Pattern Recognition:

  • Identify common experimental step sequences (e.g., "milling → sintering" for solid-state synthesis) [2]
  • Construct Markov chain representations of synthesis workflows
  • Build machine-learned flowcharts of synthesis procedures

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Computational Resources for SSL Materials Research

Resource/Tool	Function	Application in SSL Research
CGCNN Framework [4]	Crystal graph representation	Converts crystal structures to graph neural network inputs
Materials Project API [5] [4]	Data access	Provides labeled formation energy and stability data
ChemDataExtractor Toolkit [5]	Natural language processing	Automates literature data extraction from scientific manuscripts
Positive-Unlabeled Learning Algorithms [3]	Semi-supervised classification	Enables learning from positive and unlabeled examples only
Teacher-Student Dual Network [4]	Semi-supervised framework	Leverages unlabeled data through pseudo-labeling
Latent Dirichlet Allocation [2]	Topic modeling	Identifies experimental steps from synthesis text

The integration of semi-supervised learning approaches directly addresses the fundamental data scarcity problem in materials science by systematically leveraging the abundant unlabeled data that exists alongside limited labeled examples. Through frameworks including Teacher-Student Dual Neural Networks and Positive-Unlabeled learning, researchers can effectively overcome the costly absence of failed experiment data, enabling more accurate prediction of material synthesizability and stability. These methodologies represent a paradigm shift in materials informatics, transforming the high cost of failed experiments from a liability into a learning opportunity through computational approaches that extract maximum knowledge from limited experimental data.

The application of artificial intelligence in materials science, particularly for predicting material synthesizability, is often constrained by the scarcity of high-quality, labeled experimental data. While materials databases contain a wealth of structural information, most lack explicit labels for synthesizability or failed synthesis attempts. This data limitation has driven the adoption of specialized machine learning paradigms—semi-supervised, self-supervised, and positive-unlabeled (PU) learning—that can leverage both limited labeled data and abundant unlabeled data to build accurate predictive models.

Semi-supervised learning bridges supervised and unsupervised learning by using a small amount of labeled data alongside a large pool of unlabeled data. Self-supervised learning eliminates the need for manual labels altogether by creating supervisory signals directly from the structure of the data itself. Positive-unlabeled learning addresses the specific challenge where only positive examples (successfully synthesized materials) are available, with no confirmed negative examples, which is a common scenario in materials science due to publication bias favoring successful syntheses.

These approaches have demonstrated remarkable success in accelerating materials discovery. For instance, in predicting the synthesizability of inorganic materials, these methods have achieved prediction accuracies exceeding 90%, significantly outperforming traditional approaches based solely on thermodynamic stability metrics like energy above hull.

Key Algorithms and Methodological Frameworks

Positive-Unlabeled (PU) Learning

PU learning has emerged as a particularly valuable framework for synthesizability prediction because it directly addresses the fundamental data constraint in materials science: the absence of verified negative examples (failed synthesis attempts) in most databases.

Core Mathematical Framework: PU learning operates on the assumption that unlabeled data ( U ) contains both positive and negative examples, but only positive examples ( P ) are explicitly identified. The key insight is that the labeled positive set is a random sample from the true positive distribution. The probability of any example ( x ) being positive can be expressed as ( p(s=1|x) = p(s=1|y=1) \cdot p(y=1|x) ), where ( s ) indicates whether an example is labeled, and ( y ) indicates its true class.

Bagging PU Learning Approach: Mordelet et al.'s bagging PU algorithm has been successfully applied to materials synthesizability prediction. This method involves training an ensemble of classifiers where each classifier is trained on all positive examples and a bootstrap sample of the unlabeled data treated as negatives. The final prediction is an aggregation across all classifiers, which helps mitigate the false negative problem where actual positive examples in the unlabeled set are incorrectly labeled as negative [6].

Self-Supervised Learning (SSL)

Self-supervised learning creates supervisory signals from the data itself without human annotation, making it ideal for leveraging the vast amounts of unlabeled crystal structure data available in materials databases.

Crystal Twins Framework: The Crystal Twins (CT) method adapts self-supervised learning principles from computer vision to crystalline materials. It employs a twin Graph Neural Network (GNN) architecture that learns representations by forcing graph latent embeddings of augmented instances from the same crystalline system to be similar. Two primary implementations have been developed:

• CTBarlow: Uses Barlow Twins loss function to make the cross-correlation matrix of two embeddings as close as possible to the identity matrix
• CTSimSiam: Uses SimSiam loss function to maximize cosine similarity between embeddings of augmented instances [7]

Data Augmentation Strategies: For crystalline materials, effective augmentation techniques include random perturbations of atomic coordinates, atom masking, and edge masking in the crystal graph representation. These augmentations create different "views" of the same crystal structure while preserving its fundamental chemical identity [7].

Teacher-Student Dual Neural Network (TSDNN)

The TSDNN framework represents an advanced semi-supervised approach that specifically addresses the dataset bias in materials databases where most samples are stable, synthesizable materials.

Architecture: TSDNN employs a dual-network architecture with a teacher model trained using supervised signals and a student model that learns from both supervised signals and unsupervised feedback. The teacher model generates pseudo-labels for unlabeled data, which the student model then learns from, creating an iterative improvement cycle [4].

Implementation for Materials: When combined with a Crystal Graph Convolutional Neural Network (CGCNN), TSDNN has demonstrated significant improvements in formation energy prediction (10.3% accuracy improvement over baseline CGCNN) and synthesizability prediction (increasing true positive rate from 87.9% to 92.9% with 98% fewer parameters) [4].

Performance Comparison and Quantitative Benchmarks

Table 1: Performance Comparison of Learning Paradigms in Material Synthesizability Prediction

Method	Model Architecture	Accuracy/Performance	Key Advantages	Application Example
PU Learning	Semi-supervised bagging classifier	83.6% precision, 83.4% recall [3]	Addresses lack of negative samples; doesn't require confirmed unsynthesizable materials	Predicting synthesizability of inorganic material stoichiometries [3]
Self-Supervised Learning	Crystal Twins (CTBarlow/CTSimSiam) with CGCNN	17.09-36.97% improvement over supervised baselines on material property prediction [7]	Leverages abundant unlabeled data; no manual labeling required	Predicting formation energy, band gap, and other material properties [7]
Teacher-Student DNN	TSDNN with CGCNN encoder	92.9% true positive rate for synthesizability [4]	Handles dataset bias; improves screening accuracy with fewer parameters	Screening hypothetical cubic crystal materials [4]
Crystal Synthesis LLM	Fine-tuned large language models	98.6% synthesizability accuracy [8]	Exceptional generalization; predicts methods and precursors	Predicting synthesizability of arbitrary 3D crystal structures [8]

Table 2: Comparison with Traditional Synthesizability Assessment Methods

Method	Basis	Accuracy/Limitations	Computational Cost
Energy Above Hull	Thermodynamic stability	74.1% accuracy [8]; doesn't account for kinetic factors	Medium (requires DFT calculations)
Phonon Spectrum Analysis	Kinetic stability	82.2% accuracy [8]; computationally expensive	High (requires phonon calculations)
Machine Learning Approaches	Data-driven patterns	83.6-98.6% accuracy [3] [8]; requires quality training data	Low (after training)

Detailed Experimental Protocols

Protocol 1: PU Learning for Material Synthesizability Prediction

Objective: Predict the synthesizability likelihood for arbitrary elemental stoichiometries using positive-unlabeled learning.

Materials and Data Requirements:

• Positive samples: Experimentally synthesized materials from ICSD or Materials Project
• Unlabeled samples: Hypothetical materials from generative models or high-throughput computations
• Feature representation: Composition-based features or crystal graph representations

Step-by-Step Procedure:

• Data Preparation: Extract known synthesizable materials from trusted databases (e.g., ICSD) as positive examples. Collect a larger set of unlabeled candidates from hypothetical materials databases.
• Feature Extraction: Convert material compositions or structures into numerical feature representations. For composition-based features, consider stoichiometric attributes, elemental properties, and electronic structure descriptors.
• Model Training: Implement the bagging PU learning approach:
  • Train multiple classifiers, each using all positive examples and a random subset of unlabeled examples treated as negatives
  • Aggregate predictions across the classifier ensemble
  • Use weighted matrix factorization to identify reliable negative samples from unlabeled data
• Model Validation: Evaluate using held-out positive examples and manually verified synthesizable materials not used in training
• Continuous Synthesizability Mapping: Apply the trained model to generate synthesizability phase maps across compositional spaces [3] [6]

Expected Outcomes: The model should achieve approximately 83-84% precision and recall on test datasets and enable discovery of new phases, such as the demonstrated discovery of Cu₄FeV₃O₁₃ through guided exploration of quaternary oxide compositional space [3].

Protocol 2: Self-Supervised Learning with Crystal Twins

Objective: Learn meaningful representations of crystalline materials without property labels for improved downstream property prediction.

Materials and Data Requirements:

• Unlabeled crystal structures from materials databases (Materials Project, OQMD, etc.)
• Graph neural network architecture (CGCNN or similar)
• Data augmentation functions for crystal structures

Step-by-Step Procedure:

• Data Collection: Gather a large set of diverse crystal structures without property labels
• Data Augmentation: Implement three augmentation strategies:
  • Random perturbations of atomic coordinates (±0.05 Å)
  • Random atom masking (mask 5-15% of atoms)
  • Random edge masking (modify 5-15% of bond connections)
• Model Pretraining:
  • For CTBarlow: Train twin GNN encoders to make cross-correlation matrix of embeddings close to identity matrix
  • For CTSimSiam: Train encoder and predictor network to maximize cosine similarity between augmented views
• Downstream Fine-tuning: Transfer pretrained weights to supervised tasks and fine-tune on labeled datasets (formation energy, band gap, etc.)
• Evaluation: Benchmark performance against supervised baselines on MatBench datasets [7]

Expected Outcomes: Self-supervised pretraining should yield significant improvements (17-37%) over supervised baselines on various material property prediction tasks, particularly when labeled data is limited [7].

Protocol 3: Teacher-Student Framework for Formation Energy Prediction

Objective: Overcome dataset bias in materials databases where most samples have negative formation energy.

Materials and Data Requirements:

• Labeled materials with calculated formation energies
• Larger set of unlabeled crystal structures
• Teacher-student dual network architecture

Step-by-Step Procedure:

• Dataset Construction: Extract materials with known formation energies, acknowledging most will be negative. Supplement with unlabeled hypothetical structures.
• PU Learning Initialization: Apply iterative PU learning to identify likely negative samples from unlabeled data:
  • Start with positive (stable) and unlabeled samples
  • Randomly select unlabeled samples as negatives
  • Train initial classifier and repeat with different random samples
  • Select consensus negative samples for training
• Teacher-Student Training:
  • Teacher network generates pseudo-labels for unlabeled data
  • Student network trains on both labeled data and teacher's pseudo-labels
  • Iteratively refine both networks
• Screening Application: Apply trained model to screen hypothetical materials from generative models
• DFT Validation: Calculate formation energies of top candidates using DFT to verify predictions [4]

Expected Outcomes: The TSDNN model should achieve approximately 10% higher accuracy in formation energy classification compared to supervised baselines and successfully identify stable materials from hypothetical candidates, with >50% of recommended candidates validating as stable through DFT calculations [4].

Workflow Visualization

Workflow comparison of three semi-supervised approaches for material synthesizability prediction

Research Reagent Solutions

Table 3: Essential Computational Tools and Datasets for SSL in Material Synthesizability

Resource Name	Type	Function/Purpose	Access/Reference
Materials Project Database	Materials Database	Source of crystal structures and properties for training	materialsproject.org
Inorganic Crystal Structure Database (ICSD)	Experimental Database	Curated source of synthesizable materials as positive examples	FIZ Karlsruhe
Crystal Graph Convolutional Neural Network (CGCNN)	Software Framework	Graph neural network for learning material representations [7] [4]	Open-source Python package
MatBench	Benchmarking Suite	Standardized benchmarks for evaluating material property prediction [7]	matsci.org/matbench
PU Learning Algorithms	Algorithm Implementation	Methods for learning from positive and unlabeled data [3] [6]	Custom implementation based on published work
Crystal Twins Framework	SSL Implementation	Self-supervised learning for crystalline materials [7]	Open-source code from original publication
Teacher-Student DNN	Model Architecture	Semi-supervised framework for formation energy and synthesizability prediction [4]	GitHub: usccolumbia/tsdnn
Material Synthesis 2025 (MatSyn25)	Dataset	Large-scale 2D material synthesis processes for training [9]	arXiv:2510.00776

Applications and Case Studies

Discovery of Novel Quaternary Oxide Phase

A prominent success case for PU learning in materials science involved the discovery of a new Fe-Cu-V-O phase. Researchers first trained a PU learning model on known synthesizable inorganic materials from databases, then applied the model to explore the quaternary oxide compositional space comprising CuO, Fe₂O₃, and V₂O₅. The model suggested synthetically accessible stoichiometries, which guided experimental synthesis and led to the discovery of the previously unknown Cu₄FeV₃O₁₃ phase. This demonstrated the practical utility of synthesizability prediction in accelerating experimental materials discovery [3].

Screening Hypothetical Cubic Materials

The Teacher-Student DNN framework was successfully applied to screen novel stable cubic structures generated by a CubicGAN generative model. After training the TSDNN on formation energy and synthesizability prediction, researchers applied it to 1000 candidate samples generated by CubicGAN. DFT calculations validated that 512 of these recommended candidates had negative formation energies, confirming the model's effectiveness in identifying stable, synthesizable materials from hypothetical candidates. This approach demonstrates how SSL methods can significantly improve the efficiency of generative materials design pipelines [4].

Metal-Organic Framework Application Mapping

A multimodal self-supervised approach was developed to connect MOF synthesis to potential applications using only data available immediately after synthesis (PXRD patterns and chemical precursors). By pretraining on crystal structures from MOF databases in a self-supervised manner, the model learned meaningful representations that enabled accurate prediction of various properties, even with limited labeled data. This approach created a synthesis-to-application map for MOFs, providing insights into optimal material classes for diverse applications and demonstrating how SSL can bridge the gap between material synthesis and practical implementation [10].

Implementation Considerations and Challenges

Data Quality and Curation: The performance of all SSL methods heavily depends on data quality. Studies have shown significant discrepancies between text-mined datasets and manually curated data, with one analysis finding that only 15% of outliers in a text-mined solid-state reaction dataset were extracted correctly [6]. Manual curation, while labor-intensive, remains valuable for creating high-quality training data.

Representation Learning: Effective feature representation is crucial for material synthesizability prediction. Recent approaches have explored various representations including composition-based features, crystal graphs, and text-based representations like "material strings" for LLM fine-tuning [8]. The choice of representation significantly impacts model performance and generalizability.

Evaluation Challenges: Proper evaluation of synthesizability predictors remains challenging due to the inherent lack of verified negative examples. Cross-validation on known materials provides some indication of performance, but true validation requires experimental synthesis of predicted candidates, creating a costly feedback loop.

Computational Requirements: While SSL methods reduce the need for labeled data, they often require substantial computational resources for pretraining, particularly for self-supervised approaches working with large unlabeled datasets or complex model architectures like teacher-student networks.

The continued development of semi-supervised, self-supervised, and PU learning methods holds significant promise for addressing the fundamental challenge of material synthesizability prediction. As these approaches mature and integrate with experimental validation loops, they are poised to dramatically accelerate the discovery and synthesis of novel functional materials.

Why SSL? Leveraging Unlabeled Data in Materials Science

The discovery and development of new materials are fundamental to technological progress, impacting industries from energy to medicine. However, this process is often bottlenecked by the immense cost and time required for experimental synthesis and characterization. While machine learning (ML) promises to accelerate this discovery, its traditional supervised learning approaches require large volumes of accurately labeled data, which are expensive and time-consuming to acquire through experiments or high-fidelity simulations [7]. This data scarcity is particularly pronounced in materials science, where generating a single data point might involve complex synthesis procedures or computationally intensive quantum mechanical calculations.

Semi-supervised learning (SSL) emerges as a powerful solution to this fundamental challenge. SSL is a branch of machine learning that combines a small amount of labeled data with a large amount of unlabeled data to train models [11]. This approach is exceptionally valuable in domains like materials science, where unlabeled data—such as unpublished experimental results, uncharacterized synthesis procedures, or structures without property annotations—is often relatively abundant, while labeled data remains scarce and precious. The core premise of SSL is that the distribution of the unlabeled data, ( p(x) ), contains valuable information about the underlying data structure that can improve model performance, provided it is relevant to the specific task [11].

The application of SSL to materials science, particularly for predicting material synthesizability—whether a proposed material can be successfully synthesized—is transforming research methodologies. By leveraging both limited labeled datasets and vast pools of unlabeled data, SSL enables researchers to build more robust and accurate predictive models, guiding experimental efforts toward the most promising candidates and dramatically accelerating the materials development cycle.

Theoretical Foundations of SSL

SSL operates on several key assumptions about the relationship between the labeled and unlabeled data. When these assumptions hold, SSL algorithms can effectively leverage the unlabeled data to improve model performance significantly.

Core Assumptions of SSL

• Smoothness Assumption: If two data points ( x ) and ( x' ) in a high-density region are close, then their corresponding labels ( y ) and ( y' ) should be the same. This allows for label propagation from a labeled data point to nearby unlabeled points [11].
• Cluster Assumption: If data points are in the same cluster, they are likely to be of the same class. This implies that decision boundaries should not cut through high-density regions but instead lie in low-density regions [11].
• Manifold Assumption: High-dimensional data (such as crystal structures) often lie on a lower-dimensional manifold. Learning this manifold structure from unlabeled data can make the learning problem easier and improve generalization [11].

SSL Paradigms in Materials Research

Two main SSL paradigms are particularly prevalent in synthesizability research:

• Positive-Unlabeled (PU) Learning: This approach is used when only positive examples (e.g., known synthesizable materials) and unlabeled examples (a mixture of synthesizable and non-synthesizable materials) are available. PU learning algorithms estimate the proportion of positive examples in the unlabeled set to train a classifier [3].
• Self-Supervised Learning (SSL) as Pre-training: This involves designing a "pretext task" that does not require manual labels, such as predicting masked parts of a crystal structure or generating similar representations for slightly altered versions of the same material. The model learns meaningful representations from unlabeled data, which are then fine-tuned on a small set of labeled data for a specific downstream task like synthesizability classification [7].

SSL Applications in Material Synthesizability and Property Prediction

SSL techniques have been successfully applied to critical problems in materials science, demonstrating superior performance over traditional supervised methods, especially when labeled data is limited. The table below summarizes key applications and their outcomes.

Table 1: SSL Applications in Materials Science

Application Area	SSL Methodology	Key Outcome	Performance
Classifying Synthesis Procedures [2]	Latent Dirichlet Allocation (LDA) + Random Forest	Automated classification of solid-state, hydrothermal, and sol-gel synthesis from text.	~90% F1-score with >3000 training paragraphs [2].
Predicting Stoichiometry Synthesizability [3]	Positive-Unlabeled (PU) Learning	Predicts the likelihood of synthesizing inorganic materials from elemental stoichiometries.	83.4% recall and 83.6% estimated precision on test data [3].
Crystal Property Prediction [7]	Self-Supervised Learning (Barlow Twins, SimSiam)	Pre-training on unlabeled crystals improves downstream property prediction.	Up to 21.83% average improvement over supervised baseline on 5 property tasks [7].
3D Crystal Synthesizability Prediction [12]	Fine-tuned Large Language Models (LLMs)	Predicts synthesizability, synthetic method, and precursors for arbitrary 3D crystal structures.	98.6% synthesizability accuracy; >90% accuracy for method and precursor classification [12].

Detailed Experimental Protocols

This section provides detailed, reproducible methodologies for two prominent SSL approaches in materials synthesizability research.

Protocol 1: Positive-Unlabeled (PU) Learning for Synthesizability of Material Stoichiometry

This protocol is adapted from the work that predicted the synthesizability of inorganic compositions using PU learning, leading to the discovery of a new quaternary oxide phase [3].

Table 2: Key Research Reagents and Computational Tools for Protocol 1

Name	Function/Description	Source/Example
ICSD (Inorganic Crystal Structure Database)	Source of positive (synthesizable) examples.	FIZ Karlsruhe
Theoretical Databases (e.g., Materials Project)	Source of unlabeled examples (mixture of synthesizable and non-synthesizable materials).	materialsproject.org
PU Learning Algorithm	Algorithm to learn from positive and unlabeled data (e.g., non-negative risk estimator).	Custom Python implementation
Compositional Feature Vectors	Numerical representation of material stoichiometry (e.g., using elemental properties).	Matminer featurizer

Step-by-Step Workflow:

• Data Curation

  • Positive Data (( P )): Compile a set of known synthesizable materials from a trusted database such as the Inorganic Crystal Structure Database (ICSD). This set serves as your confirmed positive examples.
  • Unlabeled Data (( U )): Gather a larger set of material compositions from theoretical databases like the Materials Project. This set contains both synthesizable and non-synthesizable materials, but their labels are unknown.

• Feature Engineering

  • For each material composition in both ( P ) and ( U ), generate a numerical feature vector. Common features include:
    • Stoichiometric attributes (e.g., mean atomic number, weight).
    • Elemental property statistics (e.g., mean, range, variance of electronegativity, atomic radius).
    • Valence electron information.

• Model Training with PU Learning

  • Treat all data in ( P ) as labeled positive examples.
  • Treat all data in ( U ) as unlabeled.
  • Employ a PU learning algorithm, such as a non-negative risk estimator integrated with a classifier like Random Forest or a Support Vector Machine (SVM). The algorithm learns to differentiate the positive examples from the unlabeled set, effectively identifying the hidden negative examples within ( U ).

• Validation and Prediction

  • Validate the model on a held-out test set of known synthesizable materials to measure metrics like recall and precision.
  • Use the trained model to predict the synthesizability score for any new, arbitrary elemental stoichiometry. A continuous synthesizability phase map can be constructed to guide exploration in compositional spaces [3].

Diagram 1: PU Learning for Material Synthesizability

Protocol 2: Self-Supervised Learning for Crystal Property Prediction

This protocol is based on the "Crystal Twins" framework, which uses self-supervised pre-training on unlabeled crystal structures to boost the performance of Graph Neural Networks (GNNs) on various property prediction tasks [7].

Table 3: Key Research Reagents and Computational Tools for Protocol 2

Name	Function/Description	Source/Example
Crystal Graph Representation	Represents crystal structure as a graph (atoms=nodes, bonds=edges).	CGCNN, ALIGNN
Graph Neural Network (GNN)	Base model architecture for learning from crystal graphs.	CGCNN, GIN
SSL Framework	Framework for self-supervised pre-training (e.g., Barlow Twins, SimSiam).	Crystal Twins [7]
Unlabeled Crystal Database	Large collection of crystal structures without property labels.	Materials Project, COD

Step-by-Step Workflow:

• Data Preparation and Graph Construction

  • Collect a large dataset of unlabeled crystal structures from databases like the Materials Project.
  • Convert each crystal structure into a graph representation ( G ), where atoms are nodes and bonds are edges. Node features typically include atomic number, charge, etc.

• Data Augmentation for Crystals

  • Create two augmented views of the same crystal graph, ( G^A ) and ( G^B ). Augmentations can include:
    • Random Perturbation: Slightly randomize the coordinates of atoms within the unit cell.
    • Atom Masking: Randomly mask out a small percentage of atom nodes.
    • Edge Masking: Randomly remove a small percentage of edges.

• Self-Supervised Pre-training

  • Use a twin network architecture (e.g., Crystal Twins) with a GNN (e.g., CGCNN) as the shared encoder ( f_\theta ).
  • Pass the two augmented graphs ( G^A ) and ( G^B ) through the encoder to obtain their latent representations ( Z^A ) and ( Z^B ).
  • Minimize a self-supervised loss function to make the representations of the two augmented views similar. Two common approaches are:
    • Barlow Twins Loss: Aims to make the cross-correlation matrix of ( Z^A ) and ( Z^B ) close to the identity matrix, reducing redundancy between vector components [7].
    • SimSiam Loss: Uses a predictor network and a stop-gradient operation to prevent collapsing, maximizing the similarity between ( Z^A ) and a projected ( Z^B ) [7].

• Supervised Fine-Tuning

  • Take the pre-trained encoder ( f_\theta ) and its learned weights.
  • Replace the SSL projection head with a new task-specific head (e.g., a regression layer for predicting formation energy).
  • Fine-tune the entire model on a small, labeled dataset for a specific downstream task (e.g., predicting band gap, formation energy, or synthesizability). The model converges faster and achieves higher accuracy by leveraging the general-purpose representations learned during pre-training [7].

Diagram 2: Self-Supervised Learning for Crystals

Semi-supervised learning represents a paradigm shift in computational materials science, effectively addressing the critical bottleneck of data scarcity. By strategically leveraging the abundant unlabeled data available in materials databases, SSL enables the development of highly accurate models for predicting material synthesizability and properties with far less labeled data than required by traditional supervised methods. The outlined protocols for Positive-Unlabeled learning and Self-Supervised Learning provide a practical roadmap for researchers to integrate these powerful techniques into their workflows. As these methods continue to mature, they will play an indispensable role in accelerating the discovery and synthesis of next-generation materials, from advanced pharmaceuticals to efficient energy solutions.

Implementing SSL Frameworks for Synthesizability Prediction

Positive-Unlabeled (PU) Learning for Missing Negative Data

Predicting whether a hypothetical material can be synthesized is a critical challenge in materials science and drug development. Traditional supervised machine learning requires large, labeled datasets containing both positive examples (synthesizable materials) and negative examples (non-synthesizable materials). However, in practice, while positive examples can be obtained from databases of experimentally realized materials, reliable negative examples are exceptionally scarce because failed synthesis attempts are rarely published or systematically recorded [13] [14]. This lack of negative data creates a significant bottleneck for applying machine learning to material synthesizability prediction.

Positive-Unlabeled (PU) learning, a branch of semi-supervised learning, directly addresses this challenge. PU learning algorithms are designed to train accurate classifiers using only a set of labeled positive examples and a set of unlabeled examples (which contain a mix of both positive and unknown negative instances) [3] [13]. This paradigm is particularly well-suited for material synthesizability research, where it leverages the vast repositories of known materials as positives and uses large collections of hypothetical structures as the unlabeled set, thereby bypassing the need for explicitly labeled negative data.

Performance of PU Learning Models in Material Synthesizability Prediction

Recent research has demonstrated the effectiveness of PU learning across various material systems. The following table summarizes the performance and key attributes of several prominent models.

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

Model Name	Material System	Key Methodology	Reported Performance	Reference
SynCoTrain	Oxide crystals	Dual classifier co-training with GCNNs (SchNet & ALIGNN)	High recall on internal and leave-out test sets [14]	[14]
CSLLM (Synthesizability LLM)	Arbitrary 3D crystal structures	Fine-tuned Large Language Models on "material string" representation	98.6% accuracy [8]	[8]
SynthNN	Inorganic crystalline materials (composition-based)	Deep learning with atom2vec composition embeddings	7x higher precision than DFT formation energies; outperformed human experts [13]	[13]
Semi-Supervised Model (Jang et al.)	Inorganic materials stoichiometry	Positive-unlabeled learning on compositions	83.4% recall, 83.6% estimated precision [3]	[3]

These models consistently surpass traditional heuristic methods, such as charge-balancing or relying solely on thermodynamic stability (e.g., energy above the convex hull), which have been shown to be insufficient proxies for synthesizability [13] [14]. For instance, one study noted that more than half of the experimentally synthesized materials in the Materials Project database do not meet the charge-balancing criterion [14].

Experimental Protocol: Implementing the SynCoTrain Model

The SynCoTrain framework exemplifies a modern, robust approach to PU learning for synthesizability prediction [14]. The following is a detailed protocol for its implementation.

Data Curation and Preprocessing

• Positive Set Construction: Source crystal structures of known, synthesizable materials from experimental databases such as the Inorganic Crystal Structure Database (ICSD). For the oxide-focused SynCoTrain model, filter entries to include only oxide materials. Exclude disordered structures to focus on ordered crystal phases [8] [14].
• Unlabeled Set Construction: Compile a large set of hypothetical or computationally generated crystal structures from databases like the Materials Project (MP), the Open Quantum Materials Database (OQMD), or JARVIS. The assumption is that this set contains a mix of synthesizable and non-synthesizable materials, with the latter dominating [8].
• Data Cleaning and Standardization:
  • Filter structures based on desired criteria (e.g., maximum number of atoms per unit cell, number of distinct elements).
  • Ensure all crystal structures are in a consistent, machine-readable format, such as CIF (Crystallographic Information File) or POSCAR.
  • For graph-based models like SynCoTrain, convert crystal structures into graph representations where nodes represent atoms and edges represent atomic bonds.

Model Architecture and Training Procedure

SynCoTrain employs a co-training framework with two distinct Graph Convolutional Neural Networks (GCNNs) to mitigate model bias and enhance generalization [14].

• Classifier Selection: Implement two GCNN models with different architectural inductive biases:
  • ALIGNN (Atomistic Line Graph Neural Network): Encodes both atomic bonds and bond angles, providing a chemist's perspective [14].
  • SchNet: Uses continuous-filter convolutional layers, suited for modeling quantum interactions and providing a physicist's perspective [14].
• PU Learning Base Algorithm: Each classifier is trained using the base PU learning method by Mordelet and Vert [14]. In this approach, the model iteratively learns to distinguish the known positive examples from the unlabeled set, treating the unlabeled set as a provisional negative class while accounting for the contamination of positive examples within it.
• Co-Training Iteration:
  • Step 1: Train both ALIGNN and SchNet models independently on the labeled positive set and the current unlabeled set using the base PU learning algorithm.
  • Step 2: Each model then predicts labels for the unlabeled data.
  • Step 3: The models exchange their most confident predictions. Data points that one model classifies as highly likely to be positive are added to the other model's positive training set for the next iteration, and vice-versa for negative predictions.
  • Step 4: Repeat Steps 1-3 for a predefined number of iterations or until convergence.
• Prediction: For a new candidate material, the final synthesizability prediction is the average of the prediction scores from both the ALIGNN and SchNet models.

Model Validation

• Hold-Out Validation: Reserve a portion of the known positive examples (e.g., from ICSD) as a test set to evaluate the model's recall—its ability to correctly identify synthesizable materials [14].
• Leave-Out Validation: Test the model on a completely different set of known synthesizable materials that were not included in any part of the training process to assess generalizability [14].
• Benchmarking: Compare the model's performance against baseline methods, such as classification based on formation energy or energy above the convex hull.

Workflow Visualization: SynCoTrain for Material Synthesizability

The following diagram illustrates the iterative co-training process of the SynCoTrain model.

Successful implementation of PU learning for synthesizability prediction relies on several key computational tools and data resources.

Table 2: Essential Research Reagents for PU Learning in Material Synthesizability

Resource Name	Type	Function and Application	Reference/Availability
ICSD (Inorganic Crystal Structure Database)	Database	Primary source for labeled positive data; contains experimentally synthesized inorganic crystal structures.	[8] [13]
Materials Project (MP) Database	Database	Source for unlabeled data; contains a vast collection of computationally predicted and experimentally known structures.	[8] [14]
CIF (Crystallographic Information File)	Data Format	Standard text-based format for representing crystal structure information, including lattice parameters and atomic coordinates.	[8] [2]
ALIGNN Model	Software/Model	A Graph Neural Network that incorporates information on atomic bonds and angles for learning from crystal structures.	[14]
SchNetPack	Software/Model	A Graph Neural Network designed for learning from atomic systems using continuous-filter convolutions.	[14]
PU Learning Algorithm (Mordelet & Vert)	Algorithm	The base positive-unlabeled learning method that enables training a classifier without explicit negative examples.	[14]

Predicting whether a theoretical material can be successfully synthesized in the laboratory represents a fundamental challenge in accelerating materials discovery. Traditional approaches relying on thermodynamic stability metrics or heuristic rules face significant limitations, as they often fail to account for kinetic factors and technological constraints that fundamentally influence synthesis outcomes [15] [16]. This challenge is further compounded by a critical data scarcity problem: while validated positive examples (successfully synthesized materials) are documented in databases, explicit negative examples (failed synthesis attempts) are rarely published or systematically recorded [15] [17]. This absence of reliable negative data renders conventional supervised classification methods ineffective for the synthesizability prediction task.

Within this context, semi-supervised learning approaches, particularly Positive and Unlabeled (PU) Learning, have emerged as powerful frameworks for tackling the synthesizability prediction problem [18] [19]. SynCoTrain represents an innovative implementation of this approach, specifically designed to address the dual challenges of data scarcity and model generalization through a sophisticated dual-classifier architecture [15] [16]. By leveraging co-training principles, SynCoTrain mitigates inherent model biases while enhancing predictive reliability across diverse material systems, establishing a new paradigm for semi-supervised learning in materials informatics.

Core Methodology and Theoretical Foundation

SynCoTrain employs a co-training framework that utilizes two complementary graph convolutional neural networks (GCNNs) which iteratively exchange predictions to refine the identification of synthesizable materials from a pool of unlabeled data [16] [18]. This architecture specifically addresses the positive-unlabeled learning scenario where only confirmed synthesizable materials (positive examples) and a large set of unlabeled candidates are available, with no confirmed negative examples [15].

The theoretical foundation of SynCoTrain rests on several key principles:

• Complementary Bias Principle: Different model architectures capture distinct aspects of material representations, leading to varied inductive biases. By combining classifiers with complementary strengths, the framework reduces overall model bias and enhances generalization [16].
• Iterative Refinement: Through multiple co-training iterations, each classifier expands the positive set for the other, progressively improving the decision boundary between synthesizable and non-synthesizable materials [18].
• Bagging Ensemble Strategy: The incorporation of multiple independent models (60 runs) within each PU learner captures variability in the unlabeled data and produces more robust, averaged predictions [18].

Classifier Architectures and Material Representations

SynCoTrain leverages two distinct graph convolutional neural networks that provide complementary perspectives on material structure:

Table: SynCoTrain Dual Classifier Architectures

Classifier	Structural Representation	Architectural Approach	Representational Perspective
ALIGNN	Atomic bonds and bond angles	Line graph representation incorporating angle information	Chemist's perspective emphasizing chemical connectivity
SchNet	Continuous-filter convolutional networks	Modeling atomic interactions via continuous filters	Physicist's perspective focusing on atomic interactions

The ALIGNN (Atomistic Line Graph Neural Network) model explicitly encodes both atomic bonds and bond angles into its architectural framework, aligning closely with a chemist's intuitive understanding of molecular structure and bonding relationships [16] [18]. This approach captures intricate geometric relationships that significantly influence material stability and synthesizability.

In contrast, SchNet utilizes continuous-filter convolutional layers that model atomic interactions through learned filter functions, representing a more physics-based approach to material representation that effectively captures interatomic potentials and spatial relationships [16] [20]. This fundamental difference in representational philosophy between the two classifiers establishes the complementary relationship that SynCoTrain exploits through its co-training mechanism.

Experimental Protocols and Implementation

Data Curation and Preprocessing

The development and validation of SynCoTrain utilized oxide crystals as a case study, selected due to their extensive experimental characterization and well-documented synthesis protocols [16]. The data curation process followed these specific protocols:

• Data Source Identification: Experimental and theoretical crystal structures were obtained from the Inorganic Crystal Structure Database (ICSD) accessed through the Materials Project API [16].
• Material Filtering: The dataset was filtered to include only oxides with determinable oxidation numbers where oxygen exhibits a -2 oxidation state, using pymatgen's get_valences function [16].
• Data Cleaning: A quality control step removed approximately 1% of experimental data points with energy above hull exceeding 1eV, identified as potentially corrupt entries [16].
• Dataset Partitioning: The final curated dataset comprised 10,206 experimental (positive) structures and 31,245 theoretical (unlabeled) structures for the co-training process [16].

This careful data curation established a robust foundation for model training while ensuring chemical consistency across the material family under investigation.

Co-Training Workflow and Implementation

The SynCoTrain co-training process follows a meticulously designed iterative protocol that enables progressive refinement of synthesizability predictions:

Co-training Workflow Diagram: This visualization illustrates the iterative prediction exchange process between the two complementary classifiers.

The detailed co-training protocol consists of these critical phases:

• Initialization Phase:

  • Establish baseline PU learners for both ALIGNN and SchNet architectures
  • Initialize with confirmed positive examples (experimentally synthesized oxides)
  • Prepare unlabeled dataset (theoretical structures) for iterative processing

• Iterative Co-training Phase:

  • Step 1: ALIGNN-based PU learner predicts synthesizability scores for unlabeled data
  • Step 2: High-confidence predictions (score ≥ 0.75) are added to the positive training set
  • Step 3: SchNet-based PU learner trains on the expanded positive set and generates new predictions
  • Step 4: The process alternates between classifiers for multiple iterations (typically 2-3 cycles)
  • Step 5: A mirrored co-training series begins with SchNet as the initial classifier

• Prediction Aggregation Phase:

  • Synthesizability scores from both co-training series are averaged
  • Final binary classification applies a 0.5 probability threshold
  • Model performance is evaluated using recall metrics on internal and leave-out test sets [18]

Each base PU learner implements a bagging strategy with 60 independent runs, where random subsets of unlabeled data are treated as negative examples during each run [18]. The final synthesizability score represents the average across all runs where the specific material was excluded from training.

Model Optimization and Regularization

To prevent overfitting and enhance generalization, SynCoTrain incorporates several advanced regularization techniques in its final prediction layer:

• Label Noise Introduction: 5% of positive and negative labels are intentionally flipped to improve model robustness [20]
• Data Augmentation: Atomic position perturbations generate structural variations for training [20]
• Weighted Loss Function: A 0.45:0.55 positive-to-negative ratio in the loss function discourages over-prediction of synthesizable materials [20]
• Dropout Regularization: Implementation of 10% dropout at the embedding layer and 20% dropout at convolutional layers [20]
• Learning Rate Scheduling: Cosine annealing learning rate scheduler stabilizes training convergence [20]

These optimization strategies collectively address the challenges of training complex models on limited positive data while maintaining strong generalization performance.

Performance Evaluation and Validation

Quantitative Performance Metrics

SynCoTrain's performance was rigorously evaluated using multiple test configurations to assess both accuracy and generalizability:

Table: SynCoTrain Performance Metrics

Evaluation Metric	Description	Performance Result
Internal Test Set Recall	Model performance on held-out data from the same distribution	High recall rates (specific values not provided in search results)
Leave-out Test Set Recall	Generalization to completely excluded data partitions	High recall rates demonstrating robust generalization [16]
Final Model Accuracy	Accuracy on comprehensive test set of 5,180 samples	90.5% accuracy achieved [20]
Stability Prediction Benchmark	Comparative performance on stability prediction task	Poor performance intentional to validate PU learning reliability [16]

The model demonstrated particularly strong performance in recall metrics, essential for minimizing false negatives in synthesizability prediction [16] [18]. This high-recall performance ensures that truly synthesizable materials are correctly identified during screening processes.

Comparative Framework Analysis

SynCoTrain represents one of several emerging approaches for synthesizability prediction, each with distinct methodological frameworks:

Table: Comparative Synthesizability Prediction Approaches

Method	Learning Paradigm	Material Representation	Key Advantages
SynCoTrain	Dual-classifier PU Learning with co-training	Graph-based structural encoding (ALIGNN + SchNet)	Mitigates model bias, high recall for oxides [15] [16]
Unified Composition-Structure Model	Supervised classification with negative sampling	Composition transformer + Structure GNN ensemble	Integrates complementary signals from composition and structure [21]
Perovskite PU Learning	Single-classifier PU Learning	Compositional descriptors and DFT energies	Domain-adapted for perovskite materials [19]

This comparative analysis highlights SynCoTrain's unique contribution through its co-training architecture, specifically designed to address model bias while maintaining high performance on well-characterized material families.

Research Reagent Solutions: Computational Toolkit

Implementing the SynCoTrain framework requires specific computational tools and data resources that constitute the essential "research reagents" for reproducible synthesizability prediction:

Table: Essential Research Reagents for SynCoTrain Implementation

Resource Category	Specific Tools/Resources	Function in Research Pipeline
Data Sources	Inorganic Crystal Structure Database (ICSD), Materials Project API	Provides experimental and theoretical crystal structures for training and validation [16]
Material Analysis	pymatgen library	Determines oxidation states, performs structural analysis, and handles crystal structure data [16]
Graph Neural Networks	ALIGNN implementation, SchNetPack	Encodes crystal structures into graph representations and executes core classification algorithms [16] [18]
Validation Frameworks	Internal test sets, Leave-out test sets	Evaluates model performance and generalization capability [18]
Domain-Specific Applications	Oxide crystal databases, Perovskite datasets	Provides specialized material families for targeted synthesizability prediction [16] [19]

This computational toolkit enables researchers to implement, validate, and extend the SynCoTrain framework across diverse material systems while maintaining methodological consistency and reproducibility.

SynCoTrain establishes a robust foundation for dual-classifier co-training approaches in material synthesizability prediction, demonstrating particularly strong performance for oxide crystal systems. The framework's innovative integration of complementary GNN architectures with PU learning principles effectively addresses the critical challenges of negative data scarcity and model generalization that have historically constrained computational material discovery.

The future research trajectory for co-training models in synthesizability prediction includes several promising directions: extension to broader material families beyond oxides, integration with generative design frameworks for inverse material discovery, and incorporation of synthesis condition prediction to guide experimental realization. As semi-supervised learning methodologies continue to evolve, SynCoTrain's dual-classifier approach provides a scalable template for balancing dataset variability with computational efficiency, ultimately accelerating the discovery and deployment of novel functional materials across energy, biomedical, and electronic applications.

Graph Neural Networks for Representing Material Stoichiometry

Graph Neural Networks (GNNs) represent one of the fastest-growing classes of machine learning models with particular relevance for chemistry and materials science. They operate directly on graph or structural representations of molecules and materials, providing full access to all relevant information needed to characterize materials. In materials science, machine learning plays an increasingly important role in predicting materials properties, accelerating simulations, designing new structures, and predicting synthesis routes for new materials [22].

The fundamental advantage of GNNs stems from their ability to work directly on natural input representations of materials, which are chemical graphs of atoms and bonds, or even 3D structures or point clouds of atoms. This allows GNNs to learn internal materials representations that are informative for specific tasks such as predicting materials properties, complementing or even replacing hand-crafted feature representations traditionally used in natural sciences [22].

For stoichiometry representation specifically, GNNs offer significant advantages over compositional or fixed-sized vector representations in terms of flexibility and scalability. They can be applied to tasks requiring knowledge of functional groups, scaffolds, or the full chemical structure and its topology, making them particularly valuable for applications in drug design or materials screening [22].

Technical Foundation of GNNs for Material Representation

Graph Representation of Materials

In mathematical chemistry, graph concepts describe the structure of compounds where molecular structures are represented by undirected graphs with nodes corresponding to atoms and edges corresponding to chemical bonds. This description extends effectively to solid-state materials, though bonds might not be uniquely defined in crystals, and the exact three-dimensional arrangement of atoms plays a more decisive role [22].

The most general graph formalism defines a graph as a tuple G = (V, E) of a set of vertices v ∈ V and a set of edges e_v,w = (v, w) ∈ E, which defines connections between vertices. For materials science applications, most tasks involve graph-level predictions, particularly molecular property prediction [22].

Message Passing Framework

Most GNNs designed for chemistry and materials science can be summarized under the Message Passing Graph Neural Networks (MPNN) framework. In this approach, associated node or edge information (atom and bond types) is provided by node attributes and edge attributes. The framework involves three key phases [22]:

• Message Passing Phase: Node information is propagated as messages through edges to neighboring nodes, with each node's embedding updated based on incoming messages
• Iteration: The message passing is repeated multiple times (t = 1...K), allowing information to travel longer distances (within the K-hop neighborhood)
• Readout Phase: A graph-level embedding is obtained by pooling node embeddings of the entire graph via a parametric readout function

The mathematical formulation of the MPNN scheme is as follows [22]: $${m}{v}^{t+1}=\mathop{\sum}\limits{w\in N(v)}{M}{t}({h}{v}^{t},{h}{w}^{t},{e}{vw})$$ $${h}{v}^{t+1}={U}{t}({h}{v}^{t},{m}{v}^{t+1})$$ $$y=R({{h}_{v}^{K}| v\in G})$$

where N(v) = {u ∈ V∣(v, u) ∈ E} denotes the set of neighbors of node v, Mt(·) is the message function, Ut(·) is the node update function, and R(·) is the readout function.

GNN Architectures for Material Stoichiometry and Synthesizability

Crystal Graph Convolutional Network (CGCNet)

For representing material stoichiometry in crystalline systems, the Crystal Graph Convolutional Network (CGCNet) has demonstrated significant capabilities. This specialized GNN architecture is designed to predict properties of materials by directly working with crystal structures. In application to non-stoichiometric materials and interstitial alloys like Mo₂C and Ti₂C, CGCNet has outperformed traditional human-derived interatomic potential models (IAPs) in prediction accuracy and data efficiency [23].

A key advantage of CGCNet is its ability to extrapolate properties to larger supercells with previously unobserved atomic configurations. This capability is particularly valuable for stoichiometry representation, as it enables prediction of properties for material configurations beyond those explicitly present in the training data [23].

Explainable GNN Approaches

Understanding structure-property relationships requires explainable GNN approaches. The Crystal Graph Explainer (CGExplainer) tool has been developed to quantify the contribution of specific atomic subassemblies and their relative spatial positions to material properties. This enables systematic analysis of structure-property relationships in three-dimensional space, which is essential for interpreting how GNNs represent complex stoichiometric relationships [23].

Unlike traditional approaches that assume fixed atomic arrangements, CGExplainer can analyze models based on the relative three-dimensional positioning of atoms within crystal lattices. This capability is crucial for studying non-stoichiometric materials and solid solutions, where atoms distribute pseudo-randomly throughout the crystal lattice [23].

Adaptive Gating Mechanisms (AG-GNN)

Recent advancements in GNN architectures address challenges like over-smoothing through mechanisms such as AG-GNN's adaptive gating. This approach dynamically balances node features and graph structure through a smart switch that controls how much information flows from the graph structure versus node features at each layer. The dual-pathway design enables effective performance in both homophilic and heterophilic graphs, maintaining strong performance even with very deep architectures (up to 64 layers) where traditional GNNs typically fail due to over-smoothing [24].

Semi-Supervised Learning for Synthesizability Prediction

Positive-Unlabeled Learning Framework

Within the context of semi-supervised learning for material synthesizability research, Positive-Unlabeled (PU) learning has emerged as a powerful approach. This semi-supervised learning method is particularly valuable when only positive (successfully synthesized) and unlabeled data are available, which matches the typical scenario in materials science where failed synthesis attempts are rarely reported [3] [6].

The PU learning framework addresses the fundamental challenge in synthesizability prediction: the lack of reliable negative examples. Instead of assuming unlabeled materials are unsynthesizable, PU learning approaches treat them as a mixture of positive and negative examples, developing methods to identify likely negative instances from the unlabeled set [6].

Application to Material Synthesizability

Recent research has demonstrated successful application of PU learning to predict the synthesizability of material stoichiometry. Studies have achieved remarkable performance metrics, with true positive rates of 83.4% for test datasets and estimated precision of 83.6% [3]. This approach enables researchers to construct continuous synthesizability phase maps for arbitrary elemental combinations that align well with available synthetic data.

The practical utility of this approach was demonstrated in experimental exploration of quaternary oxide compositional space comprising CuO, Fe₂O₃, and V₂O₅, resulting in the discovery of a new phase, Cu₄FeV₃O₁₃, guided by synthesizability predictions [3].

Table 1: Performance Metrics of Synthesizability Prediction Models

Model Type	Dataset	True Positive Rate	Estimated Precision	Key Application
PU Learning Model	Ternary Oxides	83.4%	83.6%	Prediction of synthesizable stoichiometries [3]
Human-Derived IAP	Mo₂C Structures	Baseline	Baseline	Traditional approach for property prediction [23]
CGCNet GNN	Mo₂C Structures	Superior to IAP	Superior to IAP	Property extrapolation to larger supercells [23]

Quantitative Performance Comparison

Table 2: Comparative Performance of GNN Approaches for Material Property Prediction

Model Architecture	Material System	Prediction Accuracy	Data Efficiency	Extrapolation Capability
Crystal Graph Convolutional Network (CGCNet)	Mo₂C, Ti₂C	Outperforms traditional IAP models	Higher than traditional approaches	Significant improvement for larger supercells [23]
Traditional IAP Models	Mo₂C, Ti₂C	Baseline	Lower than GNNs	Limited extrapolation capability [23]
AG-GNN with Adaptive Gating	Various Graph Datasets	Up to 5.86% improvement on large networks	Maintains performance with deep layers	Resistant to over-smoothing (up to 64 layers) [24]

Experimental Protocols

Protocol 1: Crystal Graph Representation for Stoichiometric Prediction

Purpose: To create graph representations of crystalline materials for stoichiometry-based property prediction.

Materials and Software:

• Density Functional Theory (DFT) calculation software (e.g., Quantum ESPRESSO)
• Pymatgen for materials analysis
• Crystal graph conversion scripts
• GNN framework (PyTorch Geometric or similar)

Procedure:

• Dataset Generation: Construct supercells with target stoichiometry using the special quasi-random structures (SQS) algorithm to reproduce local atomic disorder arrangements [23].
• Energy Calculation: Perform first-principles calculations using DFT to determine ground-state energies of configurations.
• Graph Construction: Convert crystal structures to graph representations with nodes as atoms and edges as bonds or proximity-based connections.
• Feature Assignment: Assign node features based on atomic properties (element type, valence, etc.) and edge features based on bond characteristics.
• Model Training: Train CGCNet model using message passing framework with multiple layers (typically 3-6) to capture atomic environment information.
• Validation: Evaluate model performance on hold-out test set of crystal structures and compare against traditional IAP models.

Applications: Prediction of formation energies, stability assessment, and identification of novel synthesizable stoichiometries.

Protocol 2: PU Learning for Synthesizability Prediction

Purpose: To predict synthesizability of material compositions using positive-unlabeled learning.

Materials and Software:

• Curated dataset of synthesized materials (e.g., from ICSD, Materials Project)
• PU learning implementation (Python-based frameworks)
• Feature extraction tools for compositional descriptors
• Validation framework with known synthesized compounds

Procedure:

• Data Curation: Collect verified synthesis data from literature and databases, with careful labeling of solid-state synthesized vs. non-solid-state synthesized materials [6].
• Feature Engineering: Compute compositional and structural features for each material, including stoichiometric ratios, elemental properties, and stability metrics.
• PU Model Setup: Implement positive-unlabeled learning framework treating verified synthesized materials as positives and hypothetical materials as unlabeled.
• Model Training: Train classifier to distinguish positive examples from likely negative examples identified from the unlabeled set.
• Validation: Evaluate model using cross-validation on known synthesized materials and estimate precision using reliability metrics.
• Application: Screen hypothetical compositions from materials databases to prioritize experimental synthesis efforts.

Applications: Accelerated discovery of synthesizable materials, guidance for experimental synthesis campaigns, and construction of synthesizability phase maps.

Research Reagent Solutions

Table 3: Essential Computational Tools for GNN-Based Material Research

Tool/Resource	Type	Function	Application in Stoichiometry Research
Quantum ESPRESSO	DFT Software	First-principles electronic structure calculations	Generate training data and validate predictions [23]
Pymatgen	Materials Analysis	Python library for materials analysis	Structure manipulation, feature extraction, and dataset preparation [6]
PyTorch Geometric	GNN Framework	Deep learning on graphs	Implement CGCNet and other GNN architectures [22]
Materials Project	Database	Crystal structures and computed properties	Source of training data and hypothetical compositions [6]
ICSD	Database	Experimental crystal structure data	Source of verified synthesized materials for PU learning [6]

Workflow Visualization

GNN Workflow for Material Synthesizability Prediction

Explainable GNN Pipeline for Material Stoichiometry

Large-Scale Pre-training with Labeled and Unlabeled Data

The discovery and synthesis of new materials are fundamental to advancements in energy storage, catalysis, and drug development. However, the process is often bottlenecked by the challenge of predicting which computationally designed materials are synthesizable in the laboratory. Traditional supervised machine learning approaches for this task require large amounts of labeled data—experimentally verified synthesizable and non-synthesizable compounds—which are prohibitively expensive and time-consuming to acquire. Semi-supervised learning (SSL) presents a powerful alternative by leveraging both limited labeled data and abundant unlabeled data to build predictive models. This document details the application of SSL, particularly methods incorporating large-scale pre-training, for material synthesizability research, providing application notes and detailed experimental protocols for researchers and scientists.

Core Concepts and Assumptions of Semi-Supervised Learning

Semi-supervised learning is a branch of machine learning that combines supervised and unsupervised learning by using both labeled and unlabeled data to train models for classification and regression tasks. Its primary value lies in scenarios where obtaining sufficient labeled data is difficult or expensive, but large amounts of unlabeled data are readily available [11].

For SSL to be effective, the unlabeled data must be relevant to the specific task, and the method typically relies on one or more of the following fundamental assumptions about the data structure [11]:

• Smoothness Assumption: If two data points are close in the input space, their labels should be the same. This allows for the transitive propagation of labels from a labeled point to nearby unlabeled points.
• Cluster Assumption: Data points belonging to the same cluster (a set of points more similar to each other than to others) are likely to belong to the same class. This implies that decision boundaries should not cut through high-density data regions.
• Low-Density Assumption: The decision boundary between classes should lie in a low-density region of the input space, meaning it should not pass through areas where many data points are clustered.
• Manifold Assumption: High-dimensional data (e.g., a complex material stoichiometry) actually lies on a lower-dimensional manifold. Learning this underlying structure makes the data more separable and the learning problem more tractable.

Quantitative Performance of SSL in Materials Research

The application of SSL to materials science, particularly synthesizability prediction, has demonstrated significant promise. The following table summarizes key performance metrics from recent studies.

Table 1: Performance of SSL Models in Materials Synthesizability Prediction

Study Focus	SSL Method Used	Key Performance Metrics	Dataset Description
Synthesizability of Material Stoichiometry [3]	Positive-Unlabeled Learning	Recall: 83.4% Estimated Precision: 83.6%	Used for predicting the likelihood of synthesizing inorganic materials for any given elemental stoichiometry.
Classification of Materials Synthesis Procedures [2]	Latent Dirichlet Allocation (LDA) + Random Forest (RF)	F1 Score: ~90% (with >3000 training paragraphs) F1 Score: >80% (with a few hundred training paragraphs)	Classified synthesis paragraphs into solid-state, hydrothermal, and sol-gel methodologies from scientific text.

A comparative study exploring the limits of pre-training for image classification also provides insights relevant to representation learning. The research found that as upstream accuracy from pre-training increases, downstream task performance eventually saturates. In some cases, better downstream performance was even achieved by models with slightly lower upstream accuracy, highlighting a complex relationship between general pre-training and specific task adaptation [25].

Experimental Protocols

This section provides detailed methodologies for implementing SSL in materials research contexts.

Protocol 1: Positive-Unlabeled Learning for Synthesizability Prediction

This protocol is adapted from the study that achieved 83.4% recall in predicting material synthesizability [3].

• Data Collection and Pre-processing:

  • Labeled Data: Compile a dataset of known synthesizable material compositions (positive labels) from experimental databases.
  • Unlabeled Data: Gather a much larger set of material compositions from computational screening or literature, with no known synthesis status.
  • Feature Engineering: Represent each material composition using features derived from stoichiometry, elemental properties (e.g., electronegativity, atomic radius), and structural descriptors.

• Model Training with PU Learning:

  • Base Model Selection: Choose a probabilistic classifier (e.g., Random Forest or Neural Network) as the base learner.
  • Training Loop: Iteratively train the model to identify reliable negative examples from the unlabeled set. The model learns to distinguish confirmed synthesizable materials (positives) from the unlabeled pool, which contains both unknown positives and true negatives.
  • Stopping Criterion: Halt training when the model's performance on a held-out validation set of known positives stabilizes.

• Validation and Prediction:

  • Model Output: The model outputs a synthesizability score (between 0 and 1) for any novel material composition.
  • Experimental Guidance: Prioritize experimental synthesis efforts for compositions with the highest synthesizability scores. The protocol in [3] successfully guided the discovery of a new quaternary oxide phase, Cu4FeV3O13.

Protocol 2: Text Classification for Synthesis Procedure Extraction

This protocol outlines the semi-supervised method for classifying materials synthesis procedures from scientific text [2].

• Text Corpus Preparation:

  • Unlabeled Data: Collect a large corpus (millions of articles) of materials science literature.
  • Labeled Data: Manually annotate a small subset of paragraphs (a few hundred to a thousand per category) with synthesis labels (e.g., solid-state, hydrothermal, sol-gel, or "none").

• Unsupervised Topic Modeling (LDA):

  • Input: Process all text paragraphs, breaking them into sentences and keywords.
  • Execution: Apply Latent Dirichlet Allocation (LDA) to the entire unlabeled corpus. LDA will automatically cluster synonymous keywords into "topics" that correspond to experimental steps (e.g., "grinding," "heating," "dissolving").
  • Output: For each sentence, LDA generates a document-topic distribution vector, quantifying the prevalence of each experimental step.

• Supervised Classification (Random Forest):

  • Feature Construction: For each annotated paragraph, create a "topic n-gram" feature vector representing the sequence of LDA-derived topics in consecutive sentences.
  • Model Training: Train a Random Forest classifier on the labeled dataset using the topic n-gram features.
  • Classification: Apply the trained classifier to automatically categorize new, unlabeled synthesis paragraphs from the literature.

The workflow for this protocol is visualized below.

Comparative Framework: SSL vs. Pre-trained Models

A critical consideration is whether to leverage unlabeled data via SSL or utilize a pre-trained model (PTM). The "Few-shot SSL" framework enables a fair comparison [26].

• Problem Setup:

  • Goal: Perform a classification task (e.g., material category) with very few labeled examples.
  • SSL Path: Use the few labeled examples + a large pool of domain-specific unlabeled data.
  • PTM Path: Use the few labeled examples to fine-tune a Vision-Language Model (e.g., CLIP) pre-trained on massive web-scale datasets.

• Experimental Procedure:

  • Data Alignment: For a given task, use the same small set of labeled data for both SSL training and PTM fine-tuning.
  • Model Comparison: Evaluate both paradigms on the same test set across multiple settings (in-distribution, out-of-distribution, open-world).
  • Analysis: Determine the best approach based on data characteristics. Key findings indicate PTMs generally outperform SSL unless the data has low resolution or lacks clear semantic structure [26].

The logical relationship and decision process for this comparison are shown in the following diagram.

The Scientist's Toolkit: Research Reagent Solutions

The following table details key computational "reagents" and their functions in building SSL models for materials research.

Table 2: Essential Components for SSL in Materials Synthesizability Research

Research Reagent (Component)	Function in the Experimental Workflow	Exemplars / Notes
Labeled Data	Provides the ground truth for supervised learning, anchoring the model's predictions to known outcomes.	Small sets of experimentally verified synthesizable (and sometimes non-synthesizable) material compositions [3].
Unlabeled Data	Provides additional data structure; allows the model to learn the underlying distribution and improve generalization via SSL assumptions.	Large databases of material compositions (e.g., from high-throughput computations or unannotated literature) [3] [2].
Pre-trained Models (PTMs)	Provides a rich, generalized feature representation from large-scale pre-training, reducing dependency on large labeled datasets in the target domain.	Vision-Language Models (VLMs) like CLIP. Fine-tuning strategies include CoOp and PromptSRC [26].
Topic Modeling Algorithm	An unsupervised method to discover latent "topics" (experimental steps) from a large text corpus, creating features for classification.	Latent Dirichlet Allocation (LDA) [2].
Semi-Supervised Algorithm	The core engine that leverages both labeled and unlabeled data according to specific assumptions (smoothness, cluster, etc.).	Positive-Unlabeled Learning [3], FixMatch (consistency regularization) [26].
Feature Representation	A numerical descriptor of a material that captures its key characteristics, serving as input to the model.	Stoichiometric features, elemental properties, and for text, topic n-gram vectors [3] [2].

Semi-supervised learning represents a paradigm shift for data-driven materials research, effectively mitigating the critical bottleneck of data labeling. The protocols outlined herein—from positive-unlabeled learning for stoichiometry prediction to text mining for synthesis procedures—provide a concrete roadmap for researchers. The emerging comparison with the pretrain-finetuning paradigm offers a crucial strategic insight: while SSL remains powerful for low-resolution or semantically complex data, pre-trained models often provide a superior and more data-efficient path for tasks involving high-resolution, well-structured information. Future progress in the field will likely hinge on the deeper integration of these two approaches, such as using pre-trained knowledge to guide and enhance pseudo-labeling in semi-supervised frameworks, ultimately accelerating the discovery and synthesis of novel functional materials.

The discovery of new inorganic materials has traditionally relied on expert intuition and laborious, often serendipitous, experimental work [3]. This process presents a significant bottleneck in materials science, as the vast majority of computationally predicted candidate materials prove impractical to synthesize in the laboratory [3]. Bridging this gap between computational prediction and experimental realization is a critical challenge for accelerating materials development for applications in energy storage, catalysis, and electronic devices [3].

This application note details a case study demonstrating how semi-supervised learning, specifically positive-unlabeled (PU) learning, can guide the experimental discovery of a novel quaternary oxide, Cu₄FeV₃O₁₃. The methodology and protocols described herein were developed within the broader context of research on synthesizability prediction for inorganic crystalline materials [3] [27]. We provide a comprehensive account of the data-driven prediction model, the experimental workflow it guided, and the verification of the newly discovered phase, serving as a prototype for future materials discovery campaigns.

Synthesizability Prediction Using Semi-Supervised Learning

The Challenge of Unlabeled Data in Materials Science

A fundamental challenge in training models to predict material synthesizability is the absence of definitive negative examples. While databases like the Inorganic Crystal Structure Database (ICSD) provide a record of successfully synthesized (positive) materials, unsuccessful syntheses are rarely reported in the literature [13]. This results in a plethora of "unlabeled" materials in chemical space, the synthesizability of which is unknown.

Positive-Unlabeled (PU) Learning Model

To address this, a data-driven model based on positive-unlabeled (PU) learning was developed [3] [27]. This semi-supervised approach treats known synthesized materials from the ICSD as positive examples and all other conceivable compositions as unlabeled data. The model then learns to probabilistically reweight these unlabeled examples according to their likelihood of being synthesizable [13] [3].

• Objective: To predict, for any given elemental stoichiometry, the likelihood of synthesizing an inorganic material [3].
• Training Data: Known synthesized materials from the ICSD were used as positive examples [3].
• Learning Approach: The model leverages the entire space of synthesized inorganic chemical compositions to learn the hidden features and chemical principles (e.g., charge-balancing, chemical family relationships, ionicity) that characterize synthesizable materials, without requiring prior chemical knowledge as a direct input [13] [3].

Table 1: Performance Metrics of the Synthesizability Prediction Model.

Metric	Performance	Description
True Positive Rate (Recall)	83.4% [3]	Proportion of actual synthesizable materials correctly identified.
Estimated Precision	83.6% [3]	Proportion of model-predicted synthesizable materials that are likely to be truly synthesizable.

This model enables the construction of continuous synthesizability phase maps for arbitrary elemental combinations, providing a powerful tool for guiding exploration in uncharted compositional spaces [3].

Application: Discovery of a Novel Fe-Cu-V-O Phase

Experimental Objective and Workflow

The objective was to experimentally explore the quaternary oxide compositional space comprising CuO, Fe₂O₃, and V₂O₅ to discover new synthesizable phases [3]. The semi-supervised synthesizability model was used to prioritize the most promising stoichiometries for experimental investigation.

The following workflow diagram illustrates the integrated computational and experimental process that led to the discovery of the new phase.

Detailed Experimental Protocols

Protocol 1: Computational Screening for Synthesizability

Purpose: To identify the most synthetically accessible stoichiometry within the Cu-Fe-V-O quaternary system for experimental validation.

Procedure:

• Input Elemental System: Define the compositional search space by specifying the involved elements: Copper (Cu), Iron (Fe), Vanadium (V), and Oxygen (O) [3].
• Model Query: Input a wide range of stoichiometries within the quaternary space into the pre-trained PU learning model [3].
• Phase Map Construction: Use the model's predictions to construct a continuous synthesizability phase map. This map visualizes the predicted synthesizability likelihood across different compositional ratios [3].
• Candidate Selection: Analyze the phase map to identify stoichiometries with the highest predicted synthesizability scores. The model identified Cu₄FeV₃O₁₃ as a high-probability candidate for successful synthesis [3].

Software & Data Requirements:

• Pre-trained synthesizability prediction model (e.g., based on PU learning) [3].
• Computational resources for generating and evaluating multiple stoichiometries.

Protocol 2: Solid-State Synthesis of Cu₄FeV₃O₁₃

Purpose: To synthesize the computationally predicted candidate material, Cu₄FeV₃O₁₃, via a conventional solid-state reaction method.

Reagents:

• Copper(II) Oxide (CuO), powder, ≥99.99% trace metals basis.
• Iron(III) Oxide (Fe₂O₃), powder, ≥99.99% trace metals basis.
• Vanadium(V) Oxide (V₂O₅), powder, ≥99.99% trace metals basis.

Table 2: Research Reagent Solutions for Solid-State Synthesis.

Reagent / Material	Function in Reaction	Purity & Form
Copper(II) Oxide (CuO)	Source of Copper cations	Powder, ≥99.99%
Iron(III) Oxide (Fe₂O₃)	Source of Iron cations	Powder, ≥99.99%
Vanadium(V) Oxide (V₂O₅)	Source of Vanadium cations	Powder, ≥99.99%

Procedure:

• Precursor Weighing: Weigh out the precursor oxides in the stoichiometric molar ratio corresponding to the target composition Cu₄FeV₃O₁₃.
• Mechanical Milling: Transfer the powder mixture to a milling apparatus (e.g., a ball mill jar). Add grinding media (e.g., zirconia balls) and mill for several hours to ensure thorough homogenization of the reactants.
• Pelletization: After milling, compress the homogeneous powder into a pellet using a hydraulic press under appropriate pressure to ensure intimate inter-particle contact.
• Thermal Treatment (Calcination): Place the pellet in a suitable ceramic crucible (e.g., alumina). Heat the sample in a box furnace in an air atmosphere. Use a controlled heating ramp rate (e.g., 5°C/min) to a target calcination temperature (e.g., 600-800°C). Maintain at the target temperature for a prolonged period (e.g., 12-24 hours) to facilitate solid-state diffusion and reaction.
• Cooling: After the dwell time, allow the furnace to cool to room temperature naturally.

Safety Notes:

• Operations involving fine powder handling (weighing, milling) should be conducted in a fume hood to prevent inhalation.
• Appropriate personal protective equipment (PPE) including a lab coat, safety glasses, and gloves must be worn.
• Follow all institutional safety protocols for high-temperature furnace operations.

Protocol 3: Phase Identification and Characterization

Purpose: To verify the successful synthesis and confirm the novelty of the Cu₄FeV₃O₁₃ phase.

Procedure:

• Powder X-ray Diffraction (PXRD):
  • Gently grind a portion of the synthesized, cooled product into a fine powder using an agate mortar and pestle.
  • Mount the powder on a sample holder and load it into a powder X-ray diffractometer.
  • Collect diffraction data over a suitable 2θ range (e.g., 10° to 80°) using Cu Kα radiation.
• Data Analysis:
  • Process the raw PXRD data to assign peak positions (Bragg angles) and intensities.
  • Compare the experimental diffraction pattern against reference patterns for the known precursor oxides (CuO, Fe₂O₃, V₂O₅) and other known phases in the Cu-Fe-V-O system available in crystallographic databases (e.g., ICSD).
  • Confirm the formation of a new phase by identifying a set of diffraction peaks that do not correspond to any of the reactants or known compounds in the system.

Equipment:

• Powder X-ray Diffractometer (e.g., with Cu Kα source).
• Agate mortar and pestle.
• Crystallographic database (e.g., ICSD) for phase identification.

Results and Confirmation

The application of the synthesizability model to the Cu-Fe-V-O system successfully guided the discovery of a new phase, Cu₄FeV₃O₁₃ [3]. The key outcomes were:

• Successful Synthesis: Solid-state synthesis based on the model's recommendation yielded a distinct product [3].
• Novelty Confirmation: Powder X-ray diffraction (PXRD) analysis confirmed that the synthesized product was a new phase, as its diffraction pattern did not match any previously reported compounds in the ICSD for this quaternary system [3].

This result validates the practical utility of semi-supervised learning models for predicting material synthesizability and their capacity to directly inform and accelerate experimental materials discovery.

Table 3: Essential Research Reagents and Materials for Synthesis Guided by Predictive Models.

Item	Function / Application
High-Purity Precursor Oxides/Carbonates	Starting materials for solid-state synthesis of oxide materials. High purity is critical to avoid side reactions and impurities.
Predictive Synthesizability Model	A computational tool (e.g., based on PU learning) to assess the likelihood of a hypothetical material being synthesizable, prior to experimental investment [3].
Inorganic Crystal Structure Database (ICSD)	A comprehensive database of known inorganic crystal structures, used as a source of positive examples for model training and for verifying the novelty of synthesized phases [13].
Ball Mill / Grinding Apparatus	For the mechanical homogenization of solid precursor powders to ensure a uniform and reactive mixture for solid-state reactions.
High-Temperature Furnace	For performing calcination and sintering reactions at high temperatures (typically up to 1500°C or more) required for solid-state synthesis.
Powder X-ray Diffractometer (PXRD)	The primary tool for phase identification and confirmation of crystallinity in synthesized solid-state materials.

Navigating Practical Challenges and Performance Limits

Addressing Class Imbalance in Pre-training Datasets

In the field of material synthesizability research, a significant challenge is the inherent class imbalance in pre-training datasets. The number of known synthesizable materials (positive samples) is often vastly outnumbered by the hypothetical or non-synthesizable (negative) structures. This imbalance can severely bias machine learning models, causing them to overlook the minority class—the very materials researchers aim to discover. This Application Note outlines practical protocols and solutions for addressing this data imbalance, framed within a semi-supervised learning context.

The table below summarizes established and emerging techniques for handling class imbalance, with their reported performance in materials science applications.

Table 1: Techniques for Addressing Class Imbalance in Material Synthesizability Prediction

Technique Category	Specific Method	Reported Performance	Application Context
Algorithmic (PU Learning)	SynCoTrain (Co-training of SchNet & ALIGNN)	High recall on internal & leave-out test sets [14]	Semi-supervised synthesizability prediction for oxide crystals [14]
Data-Level (Synthetic Data)	MatWheel Framework (Conditional Generative Models)	Performance close to or exceeding real samples in data-scarce scenarios [28]	Fully-supervised and semi-supervised property prediction [28]
Data-Level (Oversampling)	SMOTE with Ensemble Models (e.g., AdaBoost)	F1-Score of 87.6% in churn prediction (analogous domain) [29]	Balancing datasets for improved model sensitivity [29] [30]
Data-Level (Undersampling)	K-Ratio Random Undersampling (K-RUS)	Moderate Imbalance Ratio (1:10) significantly enhanced model performance [31]	Prediction of anti-pathogen activity of chemical compounds [31]
Model & Evaluation	Balanced Accuracy (BAcc) Metric	More reliable performance evaluation than standard Accuracy under imbalance [32]	Recommended default metric for imbalanced classification tasks [32]
Large Language Models	Crystal Synthesis LLM (CSLLM)	98.6% accuracy in predicting synthesizability of 3D crystal structures [8]	Direct synthesizability, method, and precursor prediction [8]

Detailed Experimental Protocols

Protocol 1: PU-Learning with a Dual-Classifier Co-Training Framework (SynCoTrain)

This protocol is designed for scenarios where only a set of known synthesizable materials (positive) and a large pool of unlabeled materials are available.

1. Reagents and Data Sources

• Positive Data (P): Experimentally verified synthesizable crystals from databases like the Inorganic Crystal Structure Database (ICSD) [8].
• Unlabeled Data (U): A large collection of theoretical structures from sources like the Materials Project (MP) [14] [8].
• Base Classifiers: Two distinct Graph Convolutional Neural Networks (GCNNs) with different architectural biases, such as SchNet (physicist's perspective) and ALIGNN (chemist's perspective on bonds/angles) [14].

2. Procedure a. Initialization: Train both classifiers independently on the initial positive set P and a randomly sampled subset from the unlabeled data U. b. Iterative Co-Training: i. Each classifier predicts labels for the entire unlabeled pool U. ii. For each classifier, select the most confidently predicted positive samples from U. iii. Exchange these newly labeled samples between the two classifiers. iv. Retrain each classifier on its augmented training set, which now includes its own positive data and the positive data provided by the other classifier. c. Convergence: Repeat step (b) for a predefined number of iterations or until the set of labeled positives stabilizes. d. Final Prediction: The final synthesizability prediction is based on the average output of the two collaboratively trained classifiers [14].

Protocol 2: Synthetic Data Generation with Conditional Generative Models (MatWheel)

This protocol uses generative models to create a balanced training dataset, suitable for both fully-supervised and semi-supervised scenarios.

1. Reagents and Data Sources

• Training Data: A set of real material structures with associated properties.
• Generative Model: A conditional generative model, such as Con-CDVAE, capable of generating crystal structures conditioned on specific properties [28].
• Property Prediction Model: A structure-property prediction model, such as CGCNN [28].

2. Procedure for Semi-Supervised Learning a. Initial Model Training: Train the property prediction model on a small fraction (e.g., 10%) of the available real, labeled training data. b. Pseudo-Labeling: Use the trained model to generate pseudo-labels for the remaining unlabeled training data. c. Generative Model Training: Train the conditional generative model (e.g., Con-CDVAE) on the combined set of real labeled data and pseudo-labeled data. d. Synthetic Data Generation: Perform kernel density estimation (KDE) on the property distribution of the training set (real + pseudo-labeled). Sample from this KDE to create conditional property values, which are then fed into the generative model to produce a large synthetic dataset. e. Final Model Training: Retrain the property prediction model on a combination of the original small real dataset and the newly generated synthetic dataset [28].

Protocol 3: Optimized Random Undersampling (K-RUS) for Bioassay Data

This protocol is effective for highly imbalanced datasets, such as those from bioassays, where inactive compounds vastly outnumber active ones.

1. Reagents and Data Sources

• Imbalanced Dataset: Bioassay data from sources like PubChem, with a high imbalance ratio (e.g., 1:100) between active (minority) and inactive (majority) classes [31].
• Classifiers: Standard ML models like Random Forest (RF) or XGBoost, and Deep Learning models like Graph Neural Networks (GCNs) [31].

2. Procedure a. Baseline Evaluation: Train and evaluate chosen models on the original, imbalanced dataset using metrics like Balanced Accuracy and F1-score. b. Ratio Optimization: Systematically apply Random Undersampling (RUS) to the majority class (inactive compounds) to create datasets with progressively lower Imbalance Ratios (IRs), such as 1:50, 1:25, and 1:10. c. Model Retraining: Retrain the models on each of these resampled datasets. d. Performance Comparison: Evaluate the models on a held-out test set. External validation is crucial to assess generalization power. e. Implementation: Identify the optimal IR (e.g., 1:10) that provides the best balance between true positive and false positive rates for the task at hand [31].

Workflow Visualization

The following diagram illustrates the logical relationship and decision pathway for selecting an appropriate technique based on the data landscape and research goal.

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential Tools and Datasets for Imbalance Research in Material Synthesizability

Reagent / Resource	Type	Function in Research	Example Source / Reference
Inorganic Crystal Structure Database (ICSD)	Data Source	Provides experimentally verified, synthesizable crystal structures for positive examples.	[8]
Materials Project (MP) Database	Data Source	A comprehensive source of computed material structures, often used as a pool for unlabeled or negative data.	[28] [14] [8]
ALIGNN Model	Algorithm (GCNN)	A graph neural network that encodes bond and angle information; used as a base classifier in co-training frameworks.	[14]
Con-CDVAE Model	Algorithm (Generative)	A conditional generative model for creating synthetic crystal structures based on target properties.	[28]
Construction Zone	Software Tool	A Python package for algorithmic generation of complex nanoscale atomic structures for synthetic data.	[33]
Imbalanced-Learn (imblearn)	Software Library	A Python library offering a wide range of resampling techniques (e.g., SMOTE, NearMiss, Tomek Links).	[34] [30]
Balanced Accuracy (BAcc)	Evaluation Metric	A performance metric that averages recall per class, providing a reliable measure for imbalanced datasets.	[32]
CIF (Crystallographic Information File)	Data Format	A standard text file format for representing crystallographic information, used as input for models.	[8]

The Impact of Small and Imbalanced Training Sets on SSL Performance

In material synthesizability research, acquiring large, balanced, and labeled datasets of experimentally realized crystals remains a significant bottleneck. The process of labeling data is expensive, requiring domain knowledge and expert involvement [35]. This creates a scenario highly suited for Semi-Supervised Learning (SSL), which leverages small amounts of labeled data alongside abundant unlabeled data. However, SSL models applied to this domain must confront two major challenges: the inherently small size of the initial labeled sets and the severely imbalanced distribution between synthesizable (positive) and non-synthesizable (negative) material classes. This application note details these challenges and provides structured protocols for employing ensemble-based SSL to develop robust synthesizability prediction models.

The performance of various machine learning approaches on imbalanced datasets, including material synthesizability prediction, is summarized in the table below.

Table 1: Performance Comparison of Learning Approaches on Imbalanced Datasets

Learning Approach	Specific Method / Model	Dataset / Application Context	Key Performance Metric	Result
Ensemble Semi-Supervised	Self-training & Co-training with Naïve Bayes [35]	Splice Site Prediction (Genomic), 1:99 Imbalance Ratio	Classification Performance	Surpassed supervised ensemble baselines; Effective with <1% labeled data
Large Language Model (Supervised)	Crystal Synthesis LLM (CSLLM) [8]	3D Crystal Synthesizability Prediction (150,120 structures)	Prediction Accuracy	98.6%
Positive-Unlabeled Learning (Semi-Supervised)	PU Learning Model [8]	Screening non-synthesizable 3D crystal structures	--	Used to construct balanced dataset for CSLLM
Thermodynamic Stability (Traditional)	Energy Above Hull (≥0.1 eV/atom) [8]	Synthesizability Screening	Prediction Accuracy	74.1%
Kinetic Stability (Traditional)	Phonon Spectrum (Lowest freq. ≥ -0.1 THz) [8]	Synthesizability Screening	Prediction Accuracy	82.2%
Visual Self-Supervised	Web-SSL (DINOv2) [36]	Visual Question Answering (VQA)	--	Scales effectively with model & data size; Matches language-supervised performance

Experimental Protocols

Protocol A: Constructing a Balanced Dataset for Material Synthesizability Prediction

Objective: To create a comprehensive and balanced dataset of synthesizable and non-synthesizable crystal structures for training high-fidelity predictors [8].

Materials:

• Source of Positive Examples: Inorganic Crystal Structure Database (ICSD).
• Source of Candidate Structures: Theoretical databases (e.g., Materials Project, Computational Material Database, OQMD, JARVIS).
• Pre-screening Tool: A pre-trained Positive-Unlabeled (PU) learning model to calculate CLscore [8].

Methodology:

• Curate Positive Samples: Select confirmed synthesizable crystal structures from the ICSD. Apply filters as needed (e.g., exclude disordered structures, limit to ≤40 atoms, ≤7 elements).
• Generate Candidate Negative Samples: Pool a large number of theoretical structures from multiple computational databases.
• Calculate CLscore: Use the pre-trained PU learning model to assign a CLscore to every candidate structure. A lower CLscore indicates a higher probability of being non-synthesizable.
• Select Negative Samples: From the pool of candidate structures, select those with the lowest CLscores (e.g., CLscore <0.1) as the final set of negative examples.
• Validate and Balance: Ensure the CLscores for the positive samples are predominantly above the negative selection threshold (e.g., >98%). Balance the final dataset by selecting a similar number of positive and negative examples.
• Visualize and Characterize: Use dimensionality reduction techniques (e.g., t-SNE) to visualize the dataset coverage across crystal systems and compositions.

Protocol B: Ensemble-Based Semi-Supervised Learning for Imbalanced Data

Objective: To leverage ensembles of semi-supervised classifiers to improve prediction performance on highly imbalanced data when labeled data is scarce [35].

Materials:

• Base Classifier: A fast, computationally efficient classifier like Naïve Bayes.
• SSL Methods: Self-training and Co-training algorithms.
• Dataset: A training set with a very small labeled subset (e.g., <1% of total data) and a large unlabeled subset, with a high imbalance ratio (e.g., 1:99).

Methodology:

• Initialize Ensemble: Create multiple base classifiers. In self-training, these can be trained on different bootstrap samples of the original small labeled set. In co-training, classifiers are trained on different "views" (feature representations) of the data.
• Semi-Supervised Iteration: a. Predict: Use the current ensemble to predict labels for the unlabeled data. b. Select & Label: Identify the unlabeled instances for which the ensemble has the highest confidence in its predictions. c. Dynamic Balancing: During the selection process, dynamically balance the set of newly labeled instances to counter the underlying class imbalance. This prevents the model from being overwhelmed by majority class examples. d. Augment Training Set: Add the newly, pseudo-labeled instances to the training set. e. Retrain: Retrain the ensemble classifiers on the augmented training set.
• Ensemble Decision: Combine the predictions of the final ensemble models through voting or averaging to produce a single, robust classification.

Workflow Visualization

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential Resources for SSL-based Material Synthesizability Research

Item Name	Function / Description	Example / Specification
ICSD (Inorganic Crystal Structure Database)	Provides a trusted source of experimentally validated, synthesizable crystal structures to serve as positive training examples [8].	Database of crystal structures.
Computational Materials Databases	Sources of theoretical crystal structures used to generate candidate (and ultimately, negative) samples for model training [8].	Materials Project (MP), OQMD, JARVIS-DFT.
Pre-trained PU Learning Model	A tool for pre-screening vast pools of theoretical structures to identify high-confidence non-synthesizable examples, enabling the creation of balanced datasets [8].	Model producing CLscore for synthesizability likelihood.
Text Representation for Crystals	A simplified, reversible text format for representing crystal structure information (lattice, composition, coordinates, symmetry) suitable for fine-tuning LLMs [8].	"Material String" format.
Ensemble SSL Algorithms	Self-training and co-training algorithms that can leverage unlabeled data and incorporate dynamic balancing to handle severe class imbalance [35].	Implementations using Naïve Bayes or other base classifiers.
Large Language Models (LLMs)	Foundational models that can be fine-tuned for high-accuracy synthesizability prediction, as well as for predicting synthesis methods and precursors [8].	Models like LLaMA, fine-tuned on material strings.

Hyperparameter Tuning with Realistic Validation Set Sizes

In the field of materials science, particularly in predicting material synthesizability and drug discovery, the application of machine learning is often constrained by the limited availability of labeled experimental data. This data scarcity directly impacts a critical step in the machine learning pipeline: hyperparameter tuning. The performance and generalizability of models, including those used for classifying synthesis procedures or predicting molecular properties, are highly dependent on the proper selection of hyperparameters [2] [37]. However, when labeled data is limited, creating a validation set of sufficient size to reliably guide this tuning process becomes a significant challenge. An inadequately sized validation set can lead to high-variance performance estimates, ultimately resulting in the selection of suboptimal models and reduced predictive accuracy on truly unseen data [38] [39]. This article details practical protocols for effective hyperparameter tuning under these realistic constraints, framed within the context of semi-supervised learning for material synthesizability research.

The Validation Set Sizing Problem

The Role of the Validation Set

In machine learning, a validation set is a portion of the data used to provide an unbiased evaluation of a model fit during the hyperparameter tuning process [40]. Unlike the training set used to fit model parameters and the test set used for the final evaluation, the validation set is used to tune the model's architecture (hyperparameters), such as the number of hidden units in a neural network [40]. This separation is crucial to avoid overfitting, where a model performs well on its training data but fails to generalize.

Challenges of Limited Validation Data

The core challenge arises when the total pool of labeled data is small. Allocating a large portion to the validation set starves the model of training data, while a small validation set produces unreliable performance estimates due to high statistical uncertainty [38] [39]. This uncertainty makes it difficult to distinguish whether a set of hyperparameters is genuinely better or if its perceived superiority is a result of random chance in a small sample. In materials science, where data annotation often requires expert knowledge and costly experiments, this is a common scenario [3] [2].

Table 1: Impact of Validation Set Size on Performance Estimate Uncertainty (Binomial Model, 95% Confidence)

Validation Set Size	Observed Accuracy	Estimated Uncertainty (±)	Minimum Detectable Improvement
50	90%	±8.3%	>16.6%
100	90%	±5.9%	>11.8%
500	90%	±2.6%	>5.2%
1000	90%	±1.9%	>3.8%

The table above, based on a binomial confidence interval analysis [39], illustrates how smaller validation sets lead to wider uncertainty ranges. For instance, an accuracy of 90% on a validation set of 100 points could mean the true performance is between ~84% and ~96%. This makes it hard to confirm that a model showing 92% accuracy is truly better than one with 90%.

Core Protocols for Hyperparameter Tuning with Small Validation Sets

Protocol 1: K-Fold Cross-Validation

Cross-validation is a robust alternative to using a single, static validation set, especially when data is limited [38] [41].

Detailed Methodology:

• Data Splitting: Randomly partition the entire labeled dataset into k equally sized folds (e.g., k=5 or k=10).
• Iterative Training and Validation: For each unique set of hyperparameters: a. For i = 1 to k:
  • Designate fold i as the validation set.
  • Combine the remaining k-1 folds into a training set.
  • Train the model on this training set.
  • Evaluate the model on the validation fold i and record the performance metric (e.g., accuracy, F1-score). b. Calculate the mean performance across all k validation folds. This is the estimate for the hyperparameters' performance.
• Model Selection: Compare the mean cross-validation scores for all hyperparameter combinations and select the set with the highest average score.
• Final Model: After tuning, train the final model on the entire labeled dataset using the chosen optimal hyperparameters.

Considerations: While computationally intensive, k-fold cross-validation makes efficient use of limited data. A common choice is k=5 or k=10. Leave-One-Out Cross-Validation (LOOCV), where k equals the number of data points, is the most thorough but also the most computationally expensive [41].

Protocol 2: Bayesian Optimization for Efficient Search

When the hyperparameter search space is large, Bayesian optimization provides an efficient strategy by building a probabilistic model of the objective function (the validation score) to direct the search toward promising hyperparameters.

Detailed Methodology:

• Define Space: Define the ranges and distributions for each hyperparameter to be tuned.
• Build Surrogate Model: Use a surrogate model (e.g., a Gaussian Process) to approximate the validation score function based on evaluated hyperparameter sets.
• Acquisition Function: Use an acquisition function (e.g., Expected Improvement), which balances exploration and exploitation, to select the next most promising hyperparameters to evaluate.
• Iterate: Repeat the process of evaluating the proposed hyperparameters (via cross-validation) and updating the surrogate model for a fixed number of iterations or until performance plateaus.

This method is more efficient than grid or random search, requiring fewer evaluations to find good hyperparameters, which is crucial when each evaluation involves training a model [42].

Protocol 3: Nested Cross-Validation for Unbiased Performance Estimation

For a final, unbiased estimate of model performance after hyperparameter tuning, a nested cross-validation (or double cross-validation) protocol is recommended.

Detailed Methodology:

• Outer Loop: Split the data into k folds (e.g., k=5). This is the outer loop.
• Inner Loop: For each fold i in the outer loop: a. Hold out fold i as the test set. b. Use the remaining k-1 folds as the data for an inner loop, where you perform hyperparameter tuning (using, for instance, the cross-validation method from Protocol 1). c. Train a model on the entire k-1 folds with the best hyperparameters found in the inner loop. d. Evaluate this model on the held-out test fold i and record the score.
• Final Performance: The scores from all k outer test folds provide an unbiased estimate of the model's generalization performance, as the test data was never used in the tuning process.

This protocol is computationally very expensive but provides the most reliable performance estimate when data is scarce [40].

Application in Semi-Supervised Learning for Material Synthesizability

Semi-supervised learning (SSL) is particularly valuable in domains like materials science and drug development, where unlabeled data is abundant but labeled data is scarce [2] [4]. The protocols above are directly applicable to tuning SSL models.

For example, a study on classifying materials synthesis procedures used a semi-supervised approach combining Latent Dirichlet Allocation (LDA) with a Random Forest (RF) classifier [2]. The RF classifier's hyperparameters (e.g., the number of trees) were critical to achieving high performance. The researchers used learning curves to determine that a few hundred annotated paragraphs were sufficient for the model to converge to >80% F1-score, indicating that hyperparameter tuning could be effectively performed on a dataset of this manageable size [2].

Another study developed a Teacher-Student Dual Neural Network (TSDNN) for material synthesizability and formation energy prediction [4]. This model architecture itself has hyperparameters. The authors' use of a positive-unlabeled (PU) learning approach to generate training data highlights the data scarcity context, making efficient hyperparameter tuning protocols not just beneficial, but essential for success.

Table 2: Key Research Reagent Solutions for SSL in Materials Science

Reagent / Resource	Function in Workflow	Application Example
Labeled Synthesis Data	Acts as the ground truth for supervised training and validation. Small, high-quality sets are used for tuning.	1000 annotated paragraphs per synthesis method used to tune an RF classifier [2].
Large Unlabeled Corpus	Used for unsupervised pre-training, feature discovery, or pseudo-labeling in SSL models.	LDA applied to 2.2M articles to identify synthesis topics [2].
Positive-Unlabeled (PU) Algorithms	Enables training of classifiers using only positive and unlabeled data, mitigating the lack of negative samples.	Used to identify likely negative samples from unlabeled data for synthesizability prediction [4].
Teacher-Student Models (e.g., TSDNN)	A SSL architecture where a teacher model generates pseudo-labels for unlabeled data to train a student model, improving performance.	Improved true positive rate for synthesizability prediction from 87.9% to 92.9% [4].
Crystal Graph Convolutional Neural Networks (CGCNN)	A supervised model that learns material properties directly from crystal structures, serving as a baseline or component in SSL.	Used as a baseline regression model for formation energy prediction [4].

Workflow and Decision Pathway

The following diagram illustrates the integrated workflow for developing and validating a semi-supervised learning model for material synthesizability, incorporating the hyperparameter tuning protocols discussed.

Diagram 1: Integrated workflow for hyperparameter tuning and model validation in SSL for material synthesizability.

Load-Balancing Techniques for Large-Scale Multi-GPU Training

In the field of computational materials science, the application of semi-supervised learning (SSL) to predict material synthesizability has emerged as a powerful paradigm for accelerating the discovery of novel inorganic crystals [2] [3]. These models learn from a small amount of labeled experimental data and vast repositories of unlabeled structural information to identify synthesizable compositions with high accuracy [8]. However, training such models on massive crystallographic datasets and complex neural architectures demands immense computational resources that far exceed the capabilities of single-GPU systems. Effective load-balancing across multiple GPUs becomes indispensable for managing the memory, computational, and communication constraints inherent to distributed training, thereby enabling researchers to iterate rapidly and scale their models to tackle increasingly complex synthesizability predictions.

Background and Key Concepts

The Semi-Supervised Learning Context for Material Synthesizability

Semi-supervised learning approaches are particularly valuable in materials science where acquiring labeled experimental data is costly and time-consuming, while unlabeled data from computational databases is abundant. For synthesizability prediction, SSL models have been successfully applied to classify synthesis methodologies from scientific text and to predict the likelihood of successful synthesis for given stoichiometries [2] [3]. These models typically utilize a combination of unsupervised techniques like Latent Dirichlet Allocation (LDA) to identify experimental steps from literature, coupled with supervised classifiers like Random Forests to categorize synthesis methods [2]. More recent approaches have employed large language models (LLMs) fine-tuned on crystal structure data, achieving remarkable accuracy exceeding 98% in predicting synthesizability [8]. The computational demand of these models, especially when processing thousands of candidate structures, necessitates efficient distributed training strategies.

Parallelism Strategies for Distributed Training

Distributed training employs various parallelism strategies to partition the computational workload across multiple GPUs. The choice of strategy directly impacts load-balancing efficiency and is determined by model architecture, dataset characteristics, and hardware constraints.

Table: Multi-GPU Parallelism Strategies for Distributed Training

Strategy	Partitioning Approach	Key Advantages	Load-Balancing Considerations	Typical Use Cases
Data Parallelism [43] [44]	Replicates model on each GPU; splits data across GPUs	Easy implementation; linear scaling for small to medium models	Communication overhead increases with number of GPUs; batch size affects balance	Training models that fit on single GPU; SSL pre-training phases
Model Parallelism [43]	Splits model layers across different GPUs	Enables training of models too large for single GPU	Sequential dependencies can create GPU idle time; requires careful layer partitioning	Extremely large models (hundreds of billions of parameters)
Pipeline Parallelism [43] [44]	Splits model across GPUs with micro-batches	Higher GPU utilization than basic model parallelism	Pipeline "bubbles" can cause brief idle states; requires sophisticated scheduling	Large transformer models; deep neural networks
Tensor Parallelism [43] [44]	Splits individual tensors/operations across GPUs	Fine-grained control over memory and compute usage	Complex communication patterns; requires specialized implementation	Ultra-large models; transformer layers in LLMs

Load-Balancing Techniques: Analysis and Protocols

Communication-Computation Overlap Architectures

Advanced neural architectures can be specifically designed to facilitate better load-balancing. The Ladder Residual architecture modifies standard residual connections in transformer models to enable overlapping of communication and computation, effectively hiding the latency of inter-GPU communication [45]. In Tensor Parallelism setups, this approach has demonstrated 29% wall-clock speedup for a 70B parameter model distributed across 8 GPUs [45]. The key innovation lies in decoupling communication from computation through architectural modifications that allow forward passes to proceed without waiting for full synchronization.

Diagram 1: Communication-Computation Overlap in Ladder Residual Architecture

Dynamic Load-Balancing Through Expert Parallelism

For Mixture-of-Experts (MoE) models commonly used in large-scale SSL applications, expert parallelism provides a dynamic load-balancing mechanism by distributing different "expert" networks across GPUs [43] [44]. A gating network routes input tokens to appropriate experts, inherently balancing computational load based on input characteristics. Compression techniques like MoE-SVD and D²-MoE further enhance this balance by decomposing expert weights into shared bases and unique delta components, reducing parameter counts by 60% while maintaining model performance [45]. This is particularly valuable for materials science SSL workflows where models must process diverse crystal structures with varying complexity.

Quantitative Performance Analysis of Load-Balancing Methods

Table: Performance Characteristics of Load-Balancing Techniques

Technique	Compression/ Speedup Ratio	Memory Reduction	Accuracy Preservation	Implementation Complexity
Ladder Residual [45]	29% speedup (8 GPUs)	Not primary focus	Comparable to dense transformer	High (architectural changes)
MoE-SVD Compression [45]	60% compression	Significant	Minimal performance loss	Medium (decomposition)
D²-MoE Compression [45]	40-60% compression	Significant	>13% gains over other compressors	Medium (decomposition + pruning)
4-bit Quantization [45]	~4x model size reduction	~75% reduction	1-3% drop on workflow tasks; 10-15% drop on complex reasoning	Low (post-training)

Experimental Protocols for Load-Balancing Evaluation

Benchmarking Communication-Computation Overlap Efficiency

Objective: Quantify the effectiveness of Ladder Residual architectures in hiding communication latency during distributed training of SSL models for material synthesizability prediction.

Materials:

• Model Architecture: Transformer-based synthesizability predictor with Ladder Residual modifications [45]
• Dataset: Crystallographic information files (CIF) for 150,120 materials (70,120 synthesizable from ICSD, 80,000 non-synthesizable) [8]
• Hardware: 8-GPU cluster with high-speed interconnects (NVLink preferred)
• Software: PyTorch with DistributedDataParallel, DeepSpeed optimization library

Procedure:

• Implement Ladder Residual connections by decoupling residual pathways from main network blocks
• Configure Tensor Parallelism with model sharding across 8 GPUs
• Instrument training code to measure time spent in communication vs. computation phases
• Train for 100,000 steps with batch size 1024, recording throughput (samples/second)
• Compare against baseline without Ladder Residual modifications under identical settings
• Evaluate final model accuracy on synthesizability test set using F1 score and precision metrics

Validation Metrics:

• GPU utilization rate (%) across all devices
• Communication time as percentage of total step time
• Throughput improvement over baseline
• Synthesizability prediction accuracy on holdout test set

Evaluating Compression-Aware Load Balancing

Objective: Assess the impact of model compression techniques on load-balancing efficiency and training stability for large SSL models.

Materials:

• Base Model: Mixture-of-Experts transformer pre-trained on materials synthesis text [2]
• Compression Tools: MoE-SVD implementation for expert decomposition [45]
• Evaluation Dataset: Synthesis procedure paragraphs from materials literature [2]
• Performance Monitoring: GPU memory usage tracking, throughput measurement

Procedure:

• Apply MoE-SVD compression to decompose each expert into low-rank matrices U and V
• Implement selective decomposition based on sensitivity metrics (singular values, activation statistics)
• Share single V-matrix across experts while maintaining separate U-matrices
• Train compressed model with identical hyperparameters as original
• Measure memory footprint per GPU and inter-GPU communication volume
• Compare classification performance on synthesis methodology identification
• Analyze trade-offs between compression ratio, throughput, and task accuracy

Validation Metrics:

• Compression ratio (original size/compressed size)
• Expert diversity preservation through similarity analysis
• Memory utilization balance across GPUs
• Synthesis classification F1 score on validation set

Implementation Toolkit

Essential Software and Hardware Solutions

Table: Research Reagent Solutions for Multi-GPU Load-Balancing

Tool/Category	Specific Examples	Function in Load-Balancing	Implementation Considerations
Distributed Training Frameworks	PyTorch DDP [43] [46], DeepSpeed [43] [44], FairScale [43]	Manages gradient synchronization, memory optimization, and communication patterns	DeepSpeed ZeRO eliminates memory redundancies; DDP simplifies data parallelism
Model Compression Libraries	MoE-SVD [45], D²-MoE [45]	Reduces parameter counts and balances memory load across GPUs	Requires SVD implementation and sensitivity analysis for expert selection
Communication Optimization	NCCL [46], Ladder Residual [45]	Enhances communication-computation overlap and reduces synchronization overhead	Ladder Residual requires architectural modifications to standard transformers
Monitoring and Profiling	PyTorch Profiler, GPU utilization tools	Identifies load imbalances and communication bottlenecks	Critical for optimizing micro-batch sizes in pipeline parallelism

Integrated Workflow for Materials SSL Applications

Diagram 2: Load-Balancing Decision Framework for SSL Material Research

Effective load-balancing in multi-GPU training systems is not merely a performance optimization but an essential enabler for advancing semi-supervised learning applications in material synthesizability research. By strategically combining architectural innovations like Ladder Residual networks, dynamic partitioning through expert parallelism, and model compression techniques such as MoE-SVD, researchers can overcome the computational barriers that constrain model scale and experimental throughput. The protocols and analyses presented provide a roadmap for implementing these techniques within materials science workflows, potentially reducing training times by over 29% while maintaining the model integrity necessary for accurate synthesizability predictions. As SSL models continue to grow in complexity and importance for materials discovery, these load-balancing approaches will become increasingly critical for leveraging limited experimental data to uncover novel synthesizable materials with desirable properties.

Mitigating Model Bias through Iterative Co-training

The application of semi-supervised learning (SSL) to material synthesizability prediction presents a promising path to accelerate the discovery of novel compounds. However, models trained on class-imbalanced data, where confirmed synthesizable materials (positive labels) vastly outnumber confirmed unsynthesizable ones (negative labels), inherit a strong bias toward majority classes, degrading performance, especially for minority classes. This application note details the implementation of iterative co-training, a robust SSL paradigm, to mitigate such model bias. We present a structured protocol and data comparing the performance of co-training against standard SSL methods. The provided framework is designed to enable researchers in materials science and drug development to build more generalizable and fair predictive models for resource-intensive discovery tasks.

In material synthesizability research, obtaining a balanced set of labeled data is a fundamental challenge. While data on successfully synthesized materials (positive examples) can be sourced from databases like the Inorganic Crystal Structure Database (ICSD), data on unsuccessful attempts (negative examples) are rarely published [16] [3]. This results in a positive and unlabeled (PU) learning scenario and severe class imbalance. Training models on such data induces a classifier bias toward the majority class (synthesizable materials), which is then amplified when the model's own biased predictions are used as pseudo-labels for unlabeled data in standard self-training, a phenomenon known as "confirmation bias" [47] [48].

Iterative co-training addresses this by training multiple classifiers that learn from each other. The core idea is that by leveraging different model architectures or data views, the individual classifiers can develop diverse and uncorrelated decision boundaries. They then iteratively label the unlabeled data for each other, which helps reduce the reinforcement of initial biases and improves the model's generalization, particularly for underrepresented classes [47] [49]. This approach has been successfully applied in materials science, for instance, in the SynCoTrain model for synthesizability prediction, which uses a co-training framework to mitigate model bias and enhance generalizability [16].

Theoretical Foundations of Bias in SSL

The Problem of Bias Inheritance

In class-imbalanced semi-supervised learning (CISSL), the primary challenge is that pseudo-labels generated from unlabeled data often inherit and amplify the bias of the initial training distribution. This occurs because the model, biased toward majority classes early in training, generates pseudo-labels that are predominantly for those classes. When these biased pseudo-labels are used for subsequent training, they further degrade the quality of feature representations and reinforce an incorrect decision boundary, leading to poor generalization for minority classes [48]. This creates a feedback loop that is difficult to break.

How Co-Training Mitigates Bias

Co-training mitigates this through two key mechanisms: diversity and collaboration.

• Promoting Diversity: Unlike self-training with a single model, co-training employs multiple classifiers. Diversity between them can be achieved through different model architectures (e.g., SchNet vs. ALIGNN for graph-based data) or through different "views" of the data (e.g., different feature sets or data augmentations) [16] [47]. This diversity ensures that the classifiers make different, uncorrelated errors.
• Iterative Collaboration: In each iteration, each classifier labels a subset of the unlabeled data. The most confidently labeled samples from one classifier are then added to the training set of the other classifier(s). This process allows the classifiers to learn from each other's strengths and "correct" each other's biases, leading to a more robust collective model [50] [49].

The Co-Training Framework for Synthesizability Prediction

The following workflow visualizes the iterative co-training process adapted for material synthesizability prediction, integrating key steps such as data preparation, the co-training loop, and a criterion for stopping the process.

Experimental Protocols

Protocol 1: Implementing a Co-Training Framework with Dual Graph Neural Networks

This protocol is adapted from the SynCoTrain model for predicting the synthesizability of oxide crystals [16].

Objective: To build a robust synthesizability prediction model by mitigating bias through co-training with two distinct Graph Neural Networks (GNNs).

Materials and Data:

• Labeled Data: Experimentally confirmed synthesizable crystals from the ICSD.
• Unlabeled Data: Theoretically proposed structures from the Materials Project database.
• Preprocessing: Remove entries with energy above hull > 1 eV (potential corrupt data). Include only oxides with a determinable oxidation number and oxygen oxidation state of -2.

Procedure:

• Data Partitioning: Begin with a small set of labeled data (e.g., 10,206 experimental oxides) and a large pool of unlabeled data (e.g., 31,245 theoretical oxides).
• Classifier Initialization: Initialize two GNN classifiers with different architectures:
  • Classifier A (Physicist's perspective): SchNet, which uses continuous-filter convolutional layers suitable for encoding atomic structures.
  • Classifier B (Chemist's perspective): ALIGNN, which explicitly encodes both atomic bonds and bond angles.
• Iterative Co-Training Loop:
  • Step 1: Separately train both Classifier A and Classifier B on the current labeled dataset.
  • Step 2: Use each trained classifier to predict labels for the entire unlabeled dataset.
  • Step 3: For each classifier, select the top k most confident predictions for each class (synthesizable/unsynthesizable). The confidence is typically measured by the prediction probability.
  • Step 4: Add the selected high-confidence predictions from Classifier A to the labeled training set of Classifier B, and vice versa.
  • Step 5: Update the shared labeled dataset by pooling the newly added labels from both classifiers.
• Stopping Criterion: Iterate until the Average Confidence Difference—the average of the absolute difference in class prediction probabilities on the unlabeled data—stops increasing, indicating potential overfitting or that no more reliable samples can be labeled [50].

Protocol 2: Adaptive Bias Mitigation (ABM) within Co-Training

This protocol enhances standard co-training by incorporating a feature-aware bias mitigation mechanism, inspired by the ABM framework [48].

Objective: To explicitly correct feature representation bias during co-training, further improving pseudo-label quality for minority classes.

Procedure:

• Follow Steps 1-3 of Protocol 1 to initialize the co-training framework.
• Compute Batch-wise Feature Mean: Within each training mini-batch, compute a global contextual reference vector I by averaging the feature representations of all labeled and unlabeled samples in that batch. I = 1/(B + μB) * ( Σ f'(x_b) + Σ f'(u_b) ) where B is the batch size, μ is the relative size of the unlabeled batch, and f' is the feature extractor.
• Logit Adjustment: During classification, refine the logits (inputs to the softmax) of each sample by subtracting the logits corresponding to the batch-wise feature mean I. This acts as a class-agnostic reference to mitigate background bias.
• Dual-Branch Classification: Employ a dual-branch classification head. One branch is dedicated to learning class-balanced features, while the other maintains standard semantic consistency. The outputs are combined to produce the final prediction, fostering more robust and balanced feature learning.
• Integrate with Co-Training: Use this enhanced model architecture for both classifiers in the co-training loop. The improved pseudo-labels generated by this system lead to better model generalization.

Performance Data and Comparison

The following tables summarize the performance gains achievable by implementing co-training and specific bias mitigation strategies in semi-supervised learning scenarios, as reported in the literature.

Table 1: Performance comparison of SSL methods on imbalanced image classification benchmarks (CIFAR-10-LT, CIFAR-100-LT). Balanced accuracy (BACC) is reported. Adapted from [48].

Method	CIFAR-10-LT (γ=100)	CIFAR-100-LT (γ=100)	Notes
FixMatch (Baseline)	76.82%	45.63%	Standard SSL, suffers from bias
DARP	81.60%	48.92%	Pseudo-label refinement
CReST	83.84%	50.16%	Generative model-based
ABM (Ours)	89.59%	55.31%	Co-training + feature-aware bias mitigation

Table 2: Key performance metrics for the SynCoTrain model on oxide synthesizability prediction. Data based on [16].

Metric	Value	Description
Recall (Test Set)	High	The model correctly identified a high proportion of truly synthesizable materials.
Generalizability	Enhanced	The co-training framework was shown to mitigate model bias and improve performance on out-of-distribution data compared to a single model.

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential computational tools and data resources for implementing co-training for synthesizability prediction.

Item Name	Function/Description	Example/Source
ALIGNN Model	A Graph Neural Network that incorporates atomic bonds and angles to learn from crystal structures. Provides a "chemist's view" of the data.	Atomistic Line Graph Neural Network [16]
SchNet Model	A Graph Neural Network that uses continuous-filter convolutional layers to represent quantum interactions. Provides a "physicist's view" of the data.	SchNetPack [16]
ICSD	A critical source of labeled, experimentally synthesized crystal structures used as positive training data.	Inorganic Crystal Structure Database [16] [3]
Materials Project API	Provides access to a large repository of theoretical and experimental crystal structures, serving as a source of both labeled and unlabeled data.	Materials Project [16] [3]
PU Learning Algorithm	The base learning algorithm for Positive and Unlabeled data, used to handle the lack of explicit negative examples.	Mordelet and Vert method [16]
Stopping Criterion Script	Implements a method to determine the near-optimal stopping point for co-training without a validation set, preventing performance degradation.	Average Confidence Difference method [50]

Benchmarking SSL Against Other Learning Paradigms

In the field of material synthesizability research, a central challenge is the scarcity of high-quality, labeled experimental data required for supervised learning (SL). Self-supervised learning (SSL) has emerged as a powerful alternative to mitigate this dependency on manual annotations. This document provides a systematic comparison of SSL and SL, framing them within the context of material science applications. It offers detailed experimental protocols and application notes tailored for researchers and scientists.

Core Conceptual Comparison

Supervised Learning (SL) relies on manually curated, labeled datasets to train models. The learning process involves directly mapping input data to corresponding human-annotated labels, which is effective but often constrained by the cost, time, and expert knowledge required for data labeling, especially in specialized domains like materials synthesis [51] [52].

Self-Supervised Learning (SSL) is a paradigm that generates its own supervisory signals from the inherent structure of unlabeled data [53] [54]. It formulates pretext tasks—such as predicting a missing part of the data or determining the relationship between different data segments—to learn meaningful representations without human intervention. These pre-trained models can subsequently be fine-tuned on downstream tasks with limited labeled data [52] [55].

Table 1: Fundamental Characteristics of SSL and Supervised Learning

Feature	Supervised Learning (SL)	Self-Supervised Learning (SSL)
Label Requirement	Large volumes of high-quality manual labels [51]	No manual labels; generates pseudo-labels from data [53]
Core Learning Signal	Ground-truth annotations provided by humans [52]	Data's inherent structure (e.g., spatial, temporal context) [55] [54]
Primary Cost	Data annotation, which is expensive and time-consuming [51]	Computational power for pre-training [51] [53]
Typical Output	Task-specific predictions	Transferable representations for multiple downstream tasks [55]
Ideal Data Type	Large, balanced, labeled datasets	Large-scale unlabeled data, with smaller labeled sets for fine-tuning [56]

Performance and Applicability Analysis

The choice between SSL and SL involves critical trade-offs in data efficiency, computational demands, and performance, heavily dependent on dataset size and quality.

Quantitative Performance Comparison

Recent studies, particularly in data-scarce domains like medical imaging, provide a quantitative basis for comparing SSL and SL performance under various conditions [56].

Table 2: Performance Comparison of SSL vs. SL on Medical Imaging Tasks (Analogous to Materials Science Data Challenges)

Task / Condition	Dataset Size (Training)	Supervised Learning Performance	Self-Supervised Learning Performance	Key Takeaway
General Small Datasets	~800-1,200 images	Outperformed SSL in most experiments [56]	Lower performance than SL in this regime [56]	SL can be superior when labeled data is very limited.
Class-Imbalanced Data	Varies (imbalanced)	Significant performance degradation [56]	More robust; performance gap smaller than for SL [56]	SSL representations are less sensitive to class imbalance.
Larger Unlabeled Data + Limited Labels	Large unlabeled + small labeled set	Not applicable (requires labels)	Can match or exceed SL by leveraging unlabeled data [56] [57]	SSL excels by utilizing abundant unlabeled data.
Cardiac MRI T1 Mapping	60-second scan data	Lower repeatability (Coefficient of Variation: 12.0%) [57]	Higher repeatability (Coefficient of Variation: 6.3%) [57]	SSL can achieve superior quantitative measurement stability.

Decision Framework for Material Synthesizability Research

The following workflow provides a structured guideline for selecting a learning paradigm based on project-specific constraints and data availability.

Experimental Protocols

This section details specific methodologies for implementing SSL and SL, drawing from successful applications in scientific domains.

Self-Supervised Protocol for Text-Based Synthesis Classification

This protocol adapts methods used to classify materials synthesis procedures from scientific literature [2]. It is ideal for projects aiming to automatically extract and categorize synthesis information from vast numbers of unlabeled papers.

1. Objective: To train a model that can classify paragraphs of text into specific materials synthesis methodologies (e.g., solid-state, hydrothermal, sol-gel) using primarily unlabeled data.

2. Materials & Inputs:

• Data: A large corpus of text from scientific literature (e.g., 2+ million paragraphs) [2].
• Preprocessing: Standard NLP preprocessing (tokenization, stop-word removal).

3. Step-by-Step Procedure:

• Step 1 - Unsupervised Topic Modeling: Apply Latent Dirichlet Allocation (LDA) to the unlabeled text corpus. LDA will automatically cluster co-occurring keywords into "topics" that often correspond to experimental steps (e.g., "grinding," "heating," "dissolving") [2].
• Step 2 - Feature Extraction: For each paragraph in the corpus, use the fitted LDA model to compute a "document-topic" distribution vector. This vector quantifies the prevalence of each learned topic (experimental step) in the paragraph [2].
• Step 3 - Supervised Classifier Training:
  • Annotation: Manually label a relatively small subset of paragraphs (e.g., a few hundred to a thousand per category) with their correct synthesis methodology [2].
  • Training: Use the document-topic vectors from the labeled set as input features to train a Random Forest (RF) classifier. The RF learns to map sequences and combinations of experimental steps (topics) to the final synthesis methodology [2].
• Step 4 - Validation: Evaluate the trained RF classifier on a held-out test set of manually labeled paragraphs. Performance metrics (e.g., F1 score) can reach ~90% with sufficient training data [2].

Supervised Protocol for Direct Synthesizability Prediction

This protocol outlines a standard SL approach for a direct prediction task, such as classifying whether a given material stoichiometry is synthesizable [3].

1. Objective: To train a model that directly predicts a material's synthesizability (a binary or probabilistic output) from its structured features (e.g., elemental composition, stoichiometric ratios, features from periodic table).

2. Materials & Inputs:

• Data: A curated dataset of material compositions where each entry is labeled as "synthesizable" or "non-synthesizable." This requires significant upfront expert annotation [3].
• Features: Pre-computed feature vectors for each material composition.

3. Step-by-Step Procedure:

• Step 1 - Data Curation & Labeling: This is the most critical and costly step. Experts must label material compositions based on experimental data or literature evidence [3].
• Step 2 - Feature Engineering: Create or select relevant features for each material composition (e.g., elemental properties, ionic radii, electronegativity, average molecular weight).
• Step 3 - Model Training: Train a supervised model (e.g., Random Forest, Gradient Boosting, or Neural Network) on the labeled dataset. The model learns the complex relationships between the input features and the synthesizability label.
• Step 4 - Validation & Performance Analysis: Evaluate the model on a held-out test set. Report standard metrics (e.g., precision, recall, accuracy). One study achieved a test set recall of 83.4% and an estimated precision of 83.6% for synthesizability prediction [3].

The Scientist's Toolkit

This section lists key computational and data resources essential for conducting SSL and SL research in material science.

Table 3: Essential Research Reagents & Computational Tools

Tool / Resource	Type	Primary Function in Research	Relevance to Material Synthesis
Latent Dirichlet Allocation (LDA)	Algorithm	Unsupervised topic modeling from text [2]	Extracting experimental steps (e.g., "grinding", "sintering") from literature [2]
Random Forest (RF) Classifier	Algorithm	Supervised classification and regression [2]	Classifying synthesis methods from topic features or material properties [2]
Contrastive Learning (e.g., SimCLR, MoCo)	SSL Framework	Learning representations by contrasting similar and dissimilar data pairs [56] [54]	Creating meaningful representations of crystal structures or reaction pathways
Masked Autoencoder (MAE)	SSL Framework	Learning representations by reconstructing masked portions of input data [52]	Pre-training on unlabeled molecular structures or spectral data
Curated Material Synthesis Database	Dataset	Labeled data for supervised training and validation [3] [2]	Essential ground truth for training and evaluating synthesizability models [3]
Scientific Literature Corpus	Dataset	Unlabeled data for self-supervised pre-training [2]	Large-scale source for learning the language and patterns of synthesis [2]

In computational materials science, the efficiency of research hinges on the ability to extract meaningful insights from often limited and expensive-to-label data. The paradigm of semi-supervised learning leverages a small set of labeled data alongside a large corpus of unlabeled data, positioning it as a powerful approach for challenges like predicting material synthesizability. Central to understanding and implementing semi-supervised methods are two core concepts: Supervised Learning (SL) and Self-Supervised Learning (SSL). While the terms are sometimes conflated, they represent distinct methodologies with complementary strengths. This article clarifies the differences between SSL and SL, provides a quantitative comparison of their performance across various conditions, and offers detailed protocols for their application in materials research, particularly for predicting material properties from crystal structure data.

Defining the Paradigms: SL and SSL

Supervised Learning (SL)

Supervised Learning (SL) is the foundational paradigm where models are trained on a dataset containing input-output pairs. The model learns a mapping function from the input data (e.g., crystal structures) to known, human-annotated labels (e.g., formation energy, band gap). Its performance is heavily contingent on the availability, volume, and quality of these labeled examples [56] [58]. In material science, this often means relying on properties calculated via computationally intensive Density Functional Theory (DFT) simulations, which can be a significant bottleneck [7].

Self-Supervised Learning (SSL)

Self-Supervised Learning (SSL) is a subset of unsupervised learning designed to reduce dependence on manually curated labels. SSL creates its own pretext tasks from the inherent structure of unlabeled data. The model learns rich, general-purpose representations by solving these tasks. These pre-trained models can then be fine-tuned on specific downstream tasks (e.g., property prediction) with a limited set of labeled data, often leading to superior performance and generalization compared to training from scratch [7] [59] [58]. A prominent example in materials science is the Crystal Twins (CT) framework, which uses a twin Graph Neural Network to learn representations by ensuring that augmented views of the same crystal structure have similar latent embeddings [7].

Table 1: Core Conceptual Differences Between SL and SSL.

Aspect	Supervised Learning (SL)	Self-Supervised Learning (SSL)
Data Requirement	Large sets of labeled data.	Large sets of unlabeled data; small sets of labels for fine-tuning.
Learning Signal	Ground-truth labels provided by humans or simulations.	Pseudo-labels generated automatically from the data itself.
Primary Goal	Direct mapping from inputs to specific labeled outputs.	Learn general data representations transferable to multiple tasks.
Typical Workflow	Single-stage training on labeled data.	Two-stage: (1) Pre-training on pretext task, (2) Fine-tuning on downstream task.

Comparative Performance Analysis

The choice between SSL and SL is not absolute but depends on the specific research context. The following quantitative comparisons highlight key performance trade-offs.

Performance on Small and Imbalanced Datasets

In real-world medical and materials science applications, large, balanced datasets are the exception, not the rule. A comparative study on medical image classification tasks provides critical insights into this common scenario.

Table 2: SSL vs. SL Performance on Small/Imbalanced Medical Imaging Datasets (Mean AUC). Adapted from [56].

Task (Training Set Size)	Supervised Learning (SL)	Self-Supervised Learning (SSL)	Performance Note
Alzheimer's Diagnosis (n=771)	0.876	0.809	SL outperformed SSL.
Pneumonia Diagnosis (n=1,214)	0.952	0.918	SL outperformed SSL.
Age Prediction (n=843)	0.975	0.945	SL outperformed SSL.
Retinal Disease (n=33,484)	0.963	0.979	SSL outperformed SL with more data.

The data indicates that SL can maintain an advantage when the labeled training set is very small, even if that set is itself imbalanced [56]. SSL's performance improves relative to SL as the amount of (unlabeled) pre-training data increases, demonstrating its value in data-rich but label-poor environments.

SSL's Superior Data Efficiency in Materials Science

When ample unlabeled data is available, SSL demonstrates remarkable data efficiency by leveraging pre-training to create powerful foundational models.

Table 3: Performance of SSL vs. SL on Material Property Prediction Benchmarks. MAE reported for regression tasks; Accuracy for classification. Data from [7].

Property Prediction Task	Supervised CGCNN	CTBarlow (SSL)	CTSimSiam (SSL)	Relative Improvement
Formation Energy (eV/atom)	0.058	0.042	0.040	up to 31.0%
Band Gap (eV)	0.33	0.28	0.27	up to 18.2%
Fermi Energy (eV)	0.48	0.38	0.37	up to 22.9%
Is Metal? (Accuracy)	0.934	0.933	0.932	~0% (performance parity)

The CT framework models (CTBarlow and CTSimSiam), which use SSL pre-training, consistently outperform their supervised CGCNN counterpart across multiple challenging property prediction benchmarks [7]. The average improvement reported was 17.1% for CTBarlow and 21.8% for CTSimSiam, showcasing SSL's ability to learn more robust and generalizable representations from unlabeled crystalline structures.

Application Notes and Protocols for Material Synthesizability Research

Framed within a semi-supervised learning paradigm for material synthesizability, here are detailed protocols for implementing both SSL and SL approaches.

Experimental Protocol 1: Self-Supervised Pre-training with Crystal Twins

This protocol outlines the two-stage process for using the Crystal Twins framework [7].

1. Stage 1: Self-Supervised Pre-training

• Objective: Learn general-purpose representations of crystal structures without using property labels.
• Input: Large, unlabeled dataset of crystalline material structures (e.g., from the Materials Project, OQMD).
• Model: A twin Graph Neural Network (GNN) encoder (e.g., based on CGCNN architecture).
• Pretext Task: The model is trained to recognize different stochastically augmented views of the same crystal.
• Augmentation Techniques: Apply a combination of:
  • Random Perturbations: Slight random displacements of atomic coordinates.
  • Atom Masking: Randomly masking a subset of atom features.
  • Edge Masking: Randomly masking a subset of bonds in the crystal graph.
• Loss Function: Utilize either:
  • Barlow Twins Loss: Makes the cross-correlation matrix of the twin embeddings close to the identity matrix.
  • SimSiam Loss: Maximizes the cosine similarity between the twin embeddings using a predictor network and stop-gradient operation.
• Output: A pre-trained GNN encoder that maps crystal structures to a meaningful latent space.

2. Stage 2: Supervised Fine-Tuning for Synthesizability

• Objective: Adapt the pre-trained model to predict a specific synthesizability-related property (e.g., formation energy, thermodynamic stability).
• Input: The pre-trained encoder from Stage 1, plus a (smaller) dataset of labeled crystals for the target property.
• Procedure:
  • Initialize a new property prediction network with the pre-trained weights.
  • Add a new task-specific prediction head (e.g., a regression layer for formation energy).
  • Perform supervised training on the labeled dataset, potentially with a lower learning rate to avoid catastrophic forgetting of the pre-trained features.

Experimental Protocol 2: Standard Supervised Learning Baseline

This protocol establishes a baseline for comparison and is suitable when a sufficiently large and high-quality labeled dataset is already available.

• Objective: Train a model to directly predict a target property from crystal structures.
• Input: A single dataset of crystal structures with their corresponding property labels.
• Model: A GNN (e.g., CGCNN, ALIGNN) without any prior pre-training.
• Procedure:
  • The model is trained end-to-end.
  • The loss function (e.g., Mean Absolute Error for regression) is computed directly between the model's predictions and the ground-truth labels.
  • No separate pre-training stage is involved.
• Considerations: This approach is simpler to implement but may require more labeled data to achieve performance comparable to an SSL-based approach and is prone to overfitting on small datasets.

The Scientist's Toolkit: Key Research Reagents

Table 4: Essential Tools and Datasets for SSL/SL in Materials Informatics.

Tool/Reagent	Type	Function in Research
Crystal Graph Convolutional Neural Network (CGCNN) [7]	Software / Model	A foundational GNN architecture that represents crystals as graphs, enabling property prediction. Serves as a common backbone for both SL and SSL.
Materials Project / OQMD [7]	Database	Source of crystal structures (unlabeled data) and computed properties (labeled data) for pre-training and fine-tuning.
MatBench [7]	Benchmarking Suite	A standardized suite of tasks for fair evaluation and comparison of material property prediction models.
Crystal Twins (CT) Framework [7]	Software / Method	A specific SSL implementation (supports Barlow Twins & SimSiam) for learning material representations from unlabeled data.
Barlow Twins / SimSiam Loss [7]	Algorithm	Core SSL loss functions that enable effective pre-training by enforcing invariance to data augmentations.

The choice between Self-Supervised Learning and Supervised Learning is not a matter of which is universally better, but of which is more appropriate for a given research context. Supervised Learning excels when high-quality labeled data is abundant and readily available, providing a strong, straightforward baseline. Self-Supervised Learning shines in the more common scenario of data-rich but label-poor environments, leveraging vast unlabeled data to build powerful foundational models that can be efficiently adapted to specific tasks with minimal labels. For material synthesizability research, where obtaining definitive labels can be computationally prohibitive, integrating SSL into a semi-supervised workflow offers a compelling path toward more accurate, data-efficient, and generalizable predictive models.

The application of semi-supervised learning (SSL) to predict material synthesizability represents a paradigm shift in accelerated materials discovery. This approach addresses a fundamental challenge in computational materials science: most computationally predicted candidate materials are often impractical to synthesize in laboratory settings due to the complex nature of synthesis kinetics and technological constraints [3]. Unlike traditional supervised learning that requires extensive labeled datasets, SSL frameworks leverage both limited labeled data (known synthesizable materials) and abundant unlabeled data (hypothetical compositions) to build predictive models, effectively navigating the scarcity of negative examples (failed synthesis attempts) that are rarely published [14].

Within this context, the evaluation metrics of precision, recall, and robustness transcend mere statistical measures to become critical indicators of practical utility. Precision ensures limited experimental resources are not wasted on false positives, recall guarantees promising candidates are not overlooked, and robustness determines model reliability across diverse chemical spaces [60]. This protocol details standardized methodologies for evaluating these key metrics within SSL frameworks for material synthesizability prediction.

Quantitative Metrics for Synthesizability Prediction

The performance of SSL models for synthesizability prediction is quantitatively assessed through several key metrics, with precision and recall forming the foundational evaluation framework. These metrics are particularly crucial given the significant class imbalance and absence of explicit negative data typical in materials synthesizability datasets [14].

Table 1: Key Classification Metrics for SSL-based Synthesizability Models

Metric	Definition	Interpretation in Synthesizability Context	Target Value Range
Precision	Proportion of correctly predicted synthesizable materials among all predicted synthesizable materials	Measures how often the model's synthesis recommendations are correct; high precision minimizes resource waste on false positives	>80% [3]
Recall	Proportion of synthesizable materials correctly identified by the model	Measures the model's ability to discover all potentially synthesizable materials; high recall ensures promising candidates are not overlooked	>83% [3]
F1-Score	Harmonic mean of precision and recall	Balanced measure of model performance when both false positives and false negatives are important	>82% [3]
Specificity	Proportion of unsynthesizable materials correctly identified	Measures the model's ability to correctly reject materials that cannot be synthesized; particularly challenging without explicit negative examples	Varies by application

The application of these metrics in recent studies demonstrates their practical utility. For instance, one SSL model for predicting synthesizability of inorganic crystals achieved a recall of 83.4% and an estimated precision of 83.6% on test datasets [3]. Similarly, the SynCoTrain framework, which employs a dual-classifier co-training approach, demonstrated robust performance with high recall values on both internal and leave-out test sets for oxide crystals [14].

Diagram 1: Metric Calculation Workflow for Material Synthesizability

Robustness Evaluation Framework

Robustness in SSL models for synthesizability prediction encompasses multiple dimensions beyond simple classification accuracy, addressing challenges such as covariate shift, dataset bias, and algorithmic stability. The Matbench Discovery evaluation framework highlights the critical disconnect between thermodynamic stability calculations and actual synthesizability, emphasizing the need for prospective benchmarking that simulates real-world discovery campaigns [60].

Prospective vs. Retrospective Benchmarking

A fundamental robustness challenge arises from the disparity between retrospective performance on historical data and prospective performance in actual discovery workflows. Retrospective evaluations using random data splits often create artificially optimistic performance estimates, as they fail to account for the substantial covariate shift encountered when exploring new chemical spaces [60]. Prospective benchmarking incorporates test data generated through the intended discovery workflow, creating a more realistic assessment of model performance under actual application conditions.

Cross-Architecture Validation

Model robustness can be enhanced through cross-architecture validation frameworks such as the SynCoTrain approach, which employs two complementary graph convolutional neural networks: SchNet and ALIGNN [14]. SchNet utilizes continuous convolution filters suitable for encoding atomic structures (representing a physicist's perspective), while ALIGNN directly encodes atomic bonds and bond angles (aligning with a chemist's perspective). This architectural diversity helps mitigate individual model biases and improves generalization.

Table 2: Robustness Evaluation Metrics for SSL Synthesizability Models

Robustness Dimension	Evaluation Method	Interpretation	Acceptance Criteria
Architectural Stability	Performance variance across different model architectures (e.g., SchNet vs. ALIGNN) [14]	Consistency of predictions across different algorithmic approaches	<5% performance variance
Data Efficiency	Learning curves with varying labeled data proportions [61]	Model performance degradation with limited labeled samples	Graceful degradation (<15% recall drop at 10% labeled data)
Chemical Space Generalization	Leave-out testing on specific material families or composition spaces [14]	Ability to generalize to unseen material classes	>75% recall on novel compositions
Uncertainty Calibration	Comparison between prediction confidence and actual error rates [62]	Reliability of model's self-assessment for synthesizability predictions	Well-calibrated confidence scores

Experimental Protocols for Metric Evaluation

Precision and Recall Assessment Protocol

Objective: Quantify model performance in distinguishing synthesizable from unsynthesizable materials using standardized testing procedures.

Materials and Data Requirements:

• Labeled test set with confirmed synthesizable materials (minimum 500 compositions)
• Unlabeled data pool for semi-supervised learning (minimum 10,000 compositions)
• Positive and Unlabeled (PU) learning framework [14]

Procedure:

• Data Partitioning: Split available labeled data into training (70%), validation (15%), and test sets (15%) using temporal splitting to simulate realistic discovery scenarios [60].
• Model Training: Implement SSL framework using either:
  • Positive-unlabeled learning with iterative label propagation [3]
  • Dual-classifier co-training with SchNet and ALIGNN architectures [14]
  • Multi-mode augmentation with uncertainty-aware pseudo-labeling [61]
• Inference and Evaluation:
  • Generate synthesizability predictions on test set
  • Compute confusion matrix comparing predictions to ground truth
  • Calculate precision, recall, and F1-score using standard formulas
• Cross-Validation: Repeat procedure with 5 different random seeds, reporting mean and standard deviation of all metrics.

Expected Outcomes: A robust SSL model should achieve recall >83% and precision >80% on oxide crystal systems [3], with performance variations expected across different material families.

Robustness Stress Testing Protocol

Objective: Evaluate model stability under challenging conditions including limited labeled data, novel compositions, and architectural variations.

Procedure:

• Data Ablation Study:
  • Train models with progressively smaller labeled subsets (100%, 50%, 25%, 10% of available labeled data)
  • Evaluate performance degradation to assess data efficiency [61]
  • Compare against supervised baselines to quantify SSL benefits

• Cross-Architecture Validation:

  • Implement identical training procedures across multiple model architectures
  • For co-training frameworks, evaluate individual classifier performance alongside consensus predictions [14]
  • Quantify performance variance as robustness metric

• Prospective Testing:

  • Apply trained model to entirely new compositional spaces not represented in training data
  • Validate top predictions through experimental synthesis or high-fidelity simulation [3]
  • Calculate prospective precision as ratio of confirmed synthesizable materials to total predictions tested

Diagram 2: Robustness Stress Testing Protocol

Research Reagent Solutions

The experimental validation of SSL-based synthesizability predictions requires specialized computational tools and data resources. The following table outlines essential research reagents for implementing and evaluating SSL models for material synthesizability.

Table 3: Essential Research Reagents for SSL-based Synthesizability Prediction

Reagent / Resource	Type	Function	Example Sources
Crystal Graph Datasets	Data	Provides structured representation of atomic arrangements for GCNN models	Materials Project [14] [60], AFLOW [60], OQMD [60]
Positive-Unlabeled Learning Framework	Algorithm	Enables learning from only positive (synthesizable) and unlabeled examples	Mordelect-Vert PU Learning [14]
Graph Neural Network Architectures	Model	Encodes crystal structures into predictive features for synthesizability	SchNet [14], ALIGNN [14]
Uncertainty Quantification Tools	Algorithm	Estimates prediction reliability for experimental prioritization	Heteroscedastic Pseudo-Label Framework [62], Monte Carlo Dropout [62]
Benchmarking Platforms	Infrastructure	Standardized evaluation of model performance across tasks	Matbench Discovery [60], Open Catalyst Project [60]
Multi-mode Augmentation	Algorithm	Enhances sample completeness through mixed and random augmentation strategies	Intra-class random augmentation and inter-class mixed augmentation [61]

Advanced SSL Methodologies for Enhanced Metrics

Positive-Unlabeled Learning for Synthesizability

The fundamental challenge in synthesizability prediction—the absence of confirmed negative examples—makes Positive-Unlabeled (PU) learning particularly valuable. This approach iteratively identifies the most likely positive examples from the unlabeled data pool, gradually refining the decision boundary between synthesizable and unsynthesizable materials [14]. The PU learning paradigm aligns well with materials science reality, where confirmed synthesizable materials (positives) are documented in databases, while unsynthesizable materials (true negatives) are rarely reported.

Implementation typically follows the Mordelet and Vert approach [14], which treats each labeled positive example as a single cluster and iteratively assigns positive labels to unlabeled examples that appear similar to these clusters. This method has demonstrated effectiveness in predicting synthesizability for diverse crystal systems, achieving high recall rates while maintaining reasonable precision [3].

Uncertainty-Aware Pseudo-Labeling

Traditional pseudo-labeling approaches from semi-supervised classification face challenges in regression-oriented synthesizability prediction due to the continuous nature of the output space. Recent advances address this through uncertainty-aware pseudo-labeling frameworks that dynamically adjust pseudo-label influence based on calibrated uncertainty estimates [62].

The heteroscedastic pseudo-labeling framework approaches this through bi-level optimization that jointly minimizes empirical risk over all data while optimizing uncertainty estimates to enhance generalization on labeled data [62]. This approach effectively mitigates error propagation from incorrect pseudo-labels, a critical concern when prioritizing materials for experimental synthesis.

Multi-Mode Augmentation Strategies

Data augmentation in SSL for materials science requires specialized approaches beyond traditional image transformations. Multi-mode augmentation strategies simultaneously improve intra-class and inter-class sample completeness through combined random augmentation and mixed augmentation techniques [61].

Random augmentation enhances intra-class diversity by applying transformations to individual samples, while mixed augmentation generates synthetic examples by interpolating between different classes, effectively populating low-density regions of the feature space [61]. This dual approach addresses the fundamental challenge of limited labeled data in materials science applications.

The evaluation of precision, recall, and robustness in SSL models for material synthesizability prediction requires specialized protocols that address the unique challenges of materials science data. The frameworks and methodologies outlined in this document provide standardized approaches for assessing model performance, with particular emphasis on prospective validation and robustness testing under realistic discovery scenarios. As SSL methodologies continue to evolve, maintaining rigorous evaluation standards will be essential for translating computational predictions into experimentally confirmed materials, ultimately accelerating the discovery and development of novel functional materials for energy, electronics, and biomedical applications.

Analysis of Fine-Tuning Performance and Transferability

In the domain of materials science, accurately predicting material synthesizability—whether a theoretically proposed crystal structure can be successfully realized in a laboratory—represents a significant bottleneck in accelerating discovery. While high-throughput computational screenings routinely identify numerous candidates with promising properties, many prove non-synthesizable, creating a critical gap between theoretical prediction and experimental realization [63]. Within the broader context of semi-supervised learning for material synthesizability research, fine-tuning and transfer learning have emerged as pivotal strategies. These approaches leverage knowledge from large, source datasets to build accurate predictive models for target tasks where experimental data is scarce, such as synthesizability assessment [64] [8]. This application note provides a detailed analysis of the performance and transferability of fine-tuned machine learning models, with a specific focus on applications in material synthesizability and drug discovery.

Performance Analysis of Fine-Tuned Models

Fine-tuning pre-trained models on specific scientific tasks has consistently demonstrated superior performance compared to models trained from scratch, particularly on small target datasets. The following tables summarize key quantitative findings from recent studies.

Table 1: Performance Comparison of Fine-Tuned vs. Scratch Models on Material Properties [64]

Target Property	Scratch Model (R²)	Fine-Tuned Model (R²)	Pre-Training Property	Fine-Tuning Dataset Size
Formation Energy (FE)	0.920	0.936	Band Gap (BG)	800
Band Gap (BG)	0.572	0.609	Formation Energy (FE)	800
Band Gap (BG)	0.572	0.598	Dielectric Constant (DC)	800
Dielectric Constant (DC)	0.801	0.850	Band Gap (BG)	800

Table 2: State-of-the-Art Synthesizability Prediction Performance [8]

Model / Method	Target Task	Accuracy	Key Innovation
Crystal Synthesis LLM (CSLLM)	3D Crystal Synthesizability	98.6%	Fine-tuned LLM on comprehensive dataset
Teacher-Student DNN	3D Crystal Synthesizability	92.9%	Semi-supervised learning
PU Learning Model	3D Crystal Synthesizability	87.9%	Positive-Unlabeled learning
Thermodynamic (Eₕᵤₗₗ ≥ 0.1 eV/atom)	Synthesizability Screening	74.1%	Energy above convex hull
Kinetic (Phonon freq. ≥ -0.1 THz)	Synthesizability Screening	82.2%	Phonon spectrum analysis

The data in Table 1 shows that fine-tuning consistently enhances predictive performance across various material properties, even with modest fine-tuning dataset sizes. The pair-wise transfer learning approach yielded improvements in R² scores, reducing the mean absolute error (MAE) simultaneously [64]. Furthermore, as shown in Table 2, models specialized via fine-tuning for synthesizability prediction, such as the Crystal Synthesis Large Language Model (CSLLM), dramatically outperform traditional physics-based stability metrics, achieving state-of-the-art accuracy [8].

Experimental Protocols for Fine-Tuning

Protocol A: Multi-Property Pre-Training (MPT) Framework for GNNs

This protocol is designed for creating generalizable Graph Neural Network (GNN) models for material property prediction [64].

• Model Selection and Pre-Training Dataset Curation:

  • Select a GNN architecture capable of processing crystal structures, such as the Atomistic Line Graph Neural Network (ALIGNN).
  • Curate a large and diverse source dataset encompassing multiple material properties. A recommended dataset includes over 132,000 data points covering electronic (e.g., band gap), thermodynamic (e.g., formation energy), and mechanical (e.g., shear modulus) properties [64].

• Multi-Property Pre-Training (MPT):

  • Simultaneously pre-train the selected GNN model on all the curated properties from the source domain. This multi-task learning approach forces the model to learn robust, general-purpose feature representations of materials that are not overly specialized for a single property [64].

• Target Dataset Preparation:

  • Prepare the target dataset for the task of interest (e.g., synthesizability classification). This dataset is typically smaller. Ensure it is cleaned and standardized.

• Model Fine-Tuning:

  • Initialize the target model with the weights from the MPT model.
  • Re-train (fine-tune) the entire model or only the final layers on the smaller target dataset. The learning rate for fine-tuning is typically lower than that used for pre-training.
  • Strategy: Experiment with different fine-tuning strategies, such as freezing the initial layers of the network and only updating the weights of the later layers to prevent catastrophic forgetting [64].

• Model Validation:

  • Validate the performance of the fine-tuned model on a held-out test set from the target domain. Compare its performance against a model trained from scratch on the target dataset to quantify the improvement gained through transfer learning.

Protocol B: Fine-Tuning Large Language Models for Synthesizability and Precursor Prediction

This protocol outlines the process for adapting general-purpose LLMs to the specialized task of crystal synthesizability and precursor analysis [8].

• LLM and Dataset Selection:

  • Select an open-source LLM (e.g., LLaMA, Mistral) as the base model.
  • Construct a comprehensive and balanced dataset for fine-tuning.
    • Positive Samples: Obtain ~70,000 synthesizable crystal structures from experimental databases like the Inorganic Crystal Structure Database (ICSD). Filter for ordered structures with a manageable number of atoms and elements [8].
    • Negative Samples: Generate ~80,000 non-synthesizable examples by applying a pre-trained Positive-Unlabeled (PU) learning model to a large pool of theoretical structures (e.g., from the Materials Project) and selecting those with the lowest synthesizability scores [8].

• Text Representation of Crystal Structures:

  • Develop an efficient text-based representation for crystal structures. This involves creating a "material string" that compresses essential crystallographic information (lattice parameters, composition, atomic coordinates, space group, Wyckoff positions) into a format suitable for LLM processing, avoiding the redundancy of CIF or POSCAR files [8].

• Specialized LLM Fine-Tuning:

  • Fine-tune three separate LLMs for distinct but related tasks:
    • Synthesizability LLM: Fine-tune to classify a given material string as "synthesizable" or "non-synthesizable."
    • Method LLM: Fine-tune to classify the likely synthetic method (e.g., solid-state or solution).
    • Precursor LLM: Fine-tune to identify suitable chemical precursors for solid-state synthesis of binary and ternary compounds [8].
  • Fine-tuning Hyperparameters: Use a low learning rate (e.g., 1e-5 to 1e-6) and several epochs to adapt the LLM's internal knowledge to the specific domain without overfitting.

• Validation and Generalization Testing:

  • Assess the fine-tuned LLMs on a held-out test set.
  • For the Synthesizability LLM, perform an additional generalization test on complex experimental structures with large unit cells that are not represented in the training data to evaluate real-world robustness [8].

Workflow Visualization

The following diagrams, generated with Graphviz, illustrate the logical workflows for the experimental protocols described above.

GNN MPT Fine-Tuning Flow

Crystal Synthesis LLM Flow

The Scientist's Toolkit: Research Reagent Solutions

This section details the essential computational tools, datasets, and models that form the foundational "reagents" for conducting research in fine-tuning for synthesizability prediction.

Table 3: Key Research Reagents for Fine-Tuning in Material Synthesizability

Reagent Name / Type	Function / Application	Source / Reference
ALIGNN (GNN Architecture)	Processes crystal structures as atomic line graphs for accurate property prediction; serves as a base model for transfer learning.	[64]
Crystal Synthesis LLM (CSLLM)	A framework of fine-tuned LLMs for predicting synthesizability, synthesis methods, and precursors from text-based crystal representations.	[8]
"Material String"	A concise text representation of a crystal structure used to fine-tune LLMs, containing essential lattice, atomic, and symmetry information.	[8]
ICSD & MP Databases	Primary sources of positive (synthesizable) and theoretical (unlabeled) crystal structures for constructing balanced training datasets.	[63] [8]
Multi-Property Pre-Trained (MPT) Model	A GNN pre-trained simultaneously on diverse properties, serving as a robust starting point for fine-tuning on new tasks like synthesizability.	[64]
Positive-Unlabeled (PU) Learning	A semi-supervised technique to identify reliable negative samples (non-synthesizable structures) from a pool of unlabeled theoretical data.	[8]
Wyckoff Encode / Symmetry Analysis	A symmetry-guided method to efficiently sample promising regions of configuration space for synthesizable structures.	[63]

The discovery of new inorganic materials is fundamentally limited by the challenge of synthesizability. While computational models can predict millions of stable crystal structures, most remain impractical to synthesize in laboratory conditions [3] [65]. This application note details how semi-supervised learning (SSL) methodologies bridge this gap by leveraging both labeled and unlabeled materials data to predict synthesizable compositions and guide experimental validation. We focus specifically on validating predictions against known phases while simultaneously discovering novel materials, with emphasis on protocol implementation for research scientists.

Quantitative Performance of SSL Models

The table below summarizes performance metrics for recently developed SSL models in materials synthesizability prediction.

Table 1: Performance comparison of SSL-based synthesizability prediction models

Model Name	SSL Approach	Prediction Target	Key Performance Metrics	Reference
Stoichiometry SSL Model [3]	Positive-Unlabeled (PU) Learning	Material stoichiometry synthesizability	Recall: 83.4%, Estimated Precision: 83.6%	Matter (2024)
CLscore Model [8]	Positive-Unlabeled (PU) Learning	3D crystal structure synthesizability	Accuracy: 87.9% (3D crystals), >75% (2D MXenes)	Nature Communications (2025)
Teacher-Student Dual Network [8]	PU Learning Extension	3D crystal structure synthesizability	Accuracy: 92.9%	Nature Communications (2025)
CSLLM (Synthesizability LLM) [8]	Fine-tuned Large Language Model	3D crystal synthesizability & precursors	Accuracy: 98.6%, Precursor Prediction: 80.2%	Nature Communications (2025)
Synthesis Procedure Classifier [2]	LDA + Random Forest	Synthesis methodology from text	F1 Score: ~90%, Precision & Recall >90%	npj Computational Materials (2019)

Experimental Protocols for Validation

Protocol: Predictive Model Training via PU Learning

Purpose: To train a model that distinguishes synthesizable from non-synthesizable material compositions using partially labeled data. Principles: PU learning treats unknown materials as unlabeled data points, overcoming the limitation of exclusively labeled negative samples [3] [8].

Procedure:

• Data Curation:
  • Positive Samples: Collect 70,120 synthesizable crystal structures from the Inorganic Crystal Structure Database (ICSD). Exclude disordered structures [8].
  • Unlabeled Pool: Aggregate 1,401,562 theoretical structures from computational databases (Materials Project, OQMD, JARVIS) [8].
• Feature Engineering:
  • For stoichiometry models, use elemental composition features [3].
  • For structure-based models, calculate primary features including electron affinity, electronegativity, valence electron count, and structural parameters (e.g., square-net distance d_sq, nearest-neighbor distance d_nn) [66].
• Model Training:
  • Implement PU learning algorithm to estimate probability of synthesizability.
  • Models output a synthesizability score (e.g., CLscore) where scores <0.5 indicate non-synthesizability [8].
• Validation:
  • Use hold-out test sets from ICSD for quantitative accuracy, recall, and precision metrics [3].
  • Perform experimental validation campaigns on high-scoring candidates [3].

Protocol: Guided Exploration and Novel Phase Discovery

Purpose: To experimentally validate model predictions by targeting specific compositional spaces for novel phase discovery. Principles: SSL models generate continuous synthesizability phase maps to identify promising, previously unexplored compositions [3].

Procedure:

• Compositional Targeting:
  • Select a quaternary system (e.g., CuO-Fe$2$O$3$-V$2$O$5$) based on scientific interest [3].
  • Use the trained SSL model to compute synthesizability scores across the full compositional space.
• Phase Map Construction:
  • Generate a continuous synthesizability phase map identifying regions with high synthesizability probability [3].
• Synthesis Guidance:
  • Prioritize specific stoichiometries within high-probability regions for experimental synthesis.
  • Example: Target the composition corresponding to Cu$4$FeV$3$O$1$$3$ based on model guidance [3].
• Characterization & Validation:
  • Synthesize material using solid-state or hydrothermal methods per model suggestion.
  • Confirm novel phase formation and structure via X-ray diffraction (XRD) and other analytical techniques [3].

Workflow Visualization

Diagram 1: SSL workflow for material discovery and validation.

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential computational and experimental resources for SSL-driven materials discovery

Tool/Resource	Type	Function & Application	Access
ICSD [8]	Database	Source of experimentally verified synthesizable structures for positive training examples.	Licensed
Materials Project/OQMD/JARVIS [8]	Database	Sources of theoretical, unlabeled crystal structures for PU learning.	Open
CLscore [8]	Software Model	Pre-trained PU model for screening non-synthesizable structures from large theoretical pools.	Open Source
CSLLM Framework [8]	Software Model	Fine-tuned LLM for predicting synthesizability, synthetic methods, and precursors with high accuracy.	Open Source
LDA + Random Forest [2]	Algorithm	Classifies synthesis methodologies (e.g., solid-state, hydrothermal) from scientific text.	Code Libraries
ME-AI Framework [66]	Software Model	Gaussian Process model that incorporates expert intuition and experimental data for targeted discovery.	Open Source
FlowER [67]	Software Model	Generative AI for predicting chemically valid reaction pathways while conserving mass and electrons.	Open Source (GitHub)

Conclusion

Semi-supervised learning establishes a powerful and pragmatic framework for predicting material synthesizability, directly addressing the field's core challenge of labeled data scarcity. By effectively leveraging unlabeled data through methods like PU-learning and co-training, SSL achieves robust performance that often surpasses supervised learning in resource-constrained scenarios and demonstrates greater practical utility than some self-supervised paradigms. Key to success are strategies that handle class imbalance, enable realistic hyperparameter tuning, and utilize scalable architectures like Graph Neural Networks. Future directions point toward integrating SSL with multi-objective optimization for balancing synthesizability with target properties, incorporating physical laws into models, and developing large-scale, multi-modal pre-trained models. For biomedical research, these advances promise to accelerate the design of novel biomaterials and therapeutic agents by providing a more reliable and efficient path from computational prediction to experimental realization.

References

• 1. Closing the synthesis gap in computational materials design [https://chemrxiv.org/engage/chemrxiv/article-details/68ed4bffdfd0d042d1536798]
• 2. Semi-supervised machine-learning classification of ... [https://www.nature.com/articles/s41524-019-0204-1]
• 3. Synthesizability of materials stoichiometry using semi- ... [https://www.sciencedirect.com/science/article/pii/S2590238524002273]
• 4. Materials synthesizability and stability prediction using a semi ... [https://pubs.rsc.org/en/content/articlehtml/2023/dd/d2dd00098a]
• 5. Audacity of huge: overcoming challenges of data scarcity ... [https://www.sciencedirect.com/science/article/abs/pii/S2211339821001106]
• 6. Solid-state synthesizability predictions using positive ... [https://pubs.rsc.org/en/content/articlehtml/2025/dd/d5dd00065c]
• 7. Crystal twins: self-supervised learning for crystalline ... [https://www.nature.com/articles/s41524-022-00921-5]
• 8. Accurate prediction of synthesizability and precursors of 3D ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12264133/]
• 9. Material Synthesis 2025 (MatSyn25) Dataset for 2D Materials [https://arxiv.org/abs/2510.00776]
• 10. Connecting metal-organic framework synthesis to ... [https://www.nature.com/articles/s41467-025-60796-0]
• 11. What Is Semi-Supervised Learning? [https://www.ibm.com/think/topics/semi-supervised-learning]
• 12. Accurate prediction of synthesizability and precursors of 3D ... [https://www.nature.com/articles/s41467-025-61778-y]
• 13. Predicting the synthesizability of crystalline inorganic ... [https://www.nature.com/articles/s41524-023-01114-4]
• 14. SynCoTrain: a dual classifier PU-learning framework for ... [https://pubs.rsc.org/en/content/articlehtml/2025/dd/d4dd00394b]
• 15. SynCoTrain: a dual classifier PU-learning framework for ... [https://pubs.rsc.org/en/content/articlelanding/2025/dd/d4dd00394b]
• 16. SynCoTrain: A Dual Classifier PU-learning Framework ... [https://arxiv.org/html/2411.12011v1]
• 17. SynCoTrain: A Dual Classifier PU-learning Framework for ... [https://arxiv.org/abs/2411.12011]
• 18. [Literature Review] SynCoTrain: A Dual Classifier PU-learning ... [https://www.themoonlight.io/en/review/syncotrain-a-dual-classifier-pu-learning-framework-for-synthesizability-prediction]
• 19. Bridging the Synthesizability Gap in Perovskites by ... [https://arxiv.org/html/2510.06166v2]
• 20. SynCoTrain: A Dual Classifier PU-learning Framework for ... [https://chatpaper.com/chatpaper/fr/paper/83894]
• 21. A Synthesizability-Guided Pipeline for Materials Discovery [https://arxiv.org/html/2511.01790v1]
• 22. Graph neural networks for materials science and chemistry [https://www.nature.com/articles/s43246-022-00315-6]
• 23. Explainable GNN-derived structure–property relationships ... [https://pubs.rsc.org/en/content/articlehtml/2025/cp/d5cp02208h]
• 24. AG-GNN: Adaptive gating mechanism for robust node ... [https://www.sciencedirect.com/science/article/pii/S0020025525008862]
• 25. [2110.02095] Exploring the Limits of Large Scale Pre-training [https://arxiv.org/abs/2110.02095]
• 26. Unlabeled Data or Pre-trained Model: Rethinking Semi- ... [https://arxiv.org/html/2505.13317v1]
• 27. Synthesizability of materials stoichiometry using semi- ... [https://snu.elsevierpure.com/en/publications/synthesizability-of-materials-stoichiometry-using-semi-supervised]
• 28. MatWheel: Addressing Data Scarcity in Materials Science ... [https://arxiv.org/html/2504.09152v1]
• 29. Mitigating class imbalance in churn prediction with ... [https://www.nature.com/articles/s41598-025-01031-0]
• 30. A review of machine learning methods for imbalanced data ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12013631/]
• 31. Adjusted imbalance ratio leads to effective AI-based drug ... [https://www.nature.com/articles/s41598-025-15265-5]
• 32. Class imbalance should not throw you off balance [https://www.sciencedirect.com/science/article/pii/S1053811923004044]
• 33. A robust synthetic data generation framework for machine ... [https://www.nature.com/articles/s41524-024-01336-0]
• 34. Should You Use Imbalanced-Learn in 2025? [https://www.blog.trainindata.com/should-you-use-imbalanced-learn-in-2025/]
• 35. An empirical study of ensemble-based semi-supervised ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4565116/]
• 36. Scaling Language-Free Visual Representation Learning [https://arxiv.org/html/2504.01017v1]
• 37. Semi-supervised learning-based virtual adversarial ... [https://www.sciencedirect.com/science/article/pii/S1110016824015813]
• 38. Hyperparameter tunning if the validation set is not big ... [https://stats.stackexchange.com/questions/440358/hyperparameter-tunning-if-the-validation-set-is-not-big-enough]
• 39. How big should my validation set be? [https://www.rikvoorhaar.com/blog/validation_size]
• 40. Training, validation, and test data sets [https://en.wikipedia.org/wiki/Training,_validation,_and_test_data_sets]
• 41. Hyperparameters and Model Validation | Python Data Science ... [https://jakevdp.github.io/PythonDataScienceHandbook/05.03-hyperparameters-and-model-validation.html]
• 42. Is hyperparameter tuning on sample of dataset a bad idea? [https://stats.stackexchange.com/questions/233548/is-hyperparameter-tuning-on-sample-of-dataset-a-bad-idea]
• 43. A Practical Guide to Multi‑GPU Model Training [https://medium.com/@dhanayat.harshat/turbocharge-your-ai-training-a-practical-guide-to-multi-gpu-model-training-4bc38213bc44]
• 44. LLM Multi-GPU Training: A Guide for AI Engineers [https://towardsai.net/p/machine-learning/llm-multi-gpu-training-a-guide-for-ai-engineers]
• 45. Track: Poster Session 1 West [https://icml.cc/virtual/2025/session/50257]
• 46. Distributed Data Parallel: Speeding Up Deep Learning [https://www.acceldata.io/blog/how-distributed-data-parallel-transforms-deep-learning]
• 47. Semi-Supervised Learning with Multi-Head Co-Training [https://arxiv.org/html/2107.04795v3]
• 48. ABM: Adaptive bias mitigation for class-imbalanced semi- ... [https://www.sciencedirect.com/science/article/abs/pii/S0925231225022891]
• 49. Co-Training Unveiled: Harnessing Multifaceted Learning ... [https://medium.com/@data-overload/co-training-unveiled-harnessing-multifaceted-learning-for-semisupervised-success-3952de32e22d]
• 50. How and when to stop the co-training process [https://www.sciencedirect.com/science/article/abs/pii/S0957417421012021]
• 51. Self-Supervised Learning: Definition, Tutorial & Examples [https://www.v7labs.com/blog/self-supervised-learning-guide]
• 52. A survey on self-supervised learning: Recent advances ... [https://www.sciencedirect.com/science/article/abs/pii/S0925231225020818]
• 53. Self-supervised Learning Explained [https://encord.com/blog/self-supervised-learning/]
• 54. Self-Supervised Learning (SSL) [https://www.emergentmind.com/topics/self-supervised-learning-ssl]
• 55. Self-Supervised Representation Learning - Lil'Log [https://lilianweng.github.io/posts/2019-11-10-self-supervised/]
• 56. Comparative analysis of supervised and self ... [https://www.nature.com/articles/s41598-025-99000-0]
• 57. Repeatability-encouraging self-supervised learning ... - PubMed [https://pubmed.ncbi.nlm.nih.gov/40014485/]
• 58. Consequential Advancements of Self-Supervised Learning ... [https://www.mdpi.com/2227-7390/12/5/758]
• 59. Leveraging unlabeled SEM datasets with self-supervised ... [https://www.nature.com/articles/s41524-025-01802-3]
• 60. A framework to evaluate machine learning crystal stability ... [https://www.nature.com/articles/s42256-025-01055-1]
• 61. A new method of semi-supervised learning classification ... [https://www.nature.com/articles/s41598-025-02324-0]
• 62. Semi-Supervised Regression with Heteroscedastic Pseudo ... [https://arxiv.org/html/2510.15266v1]
• 63. Machine-Learning-Assisted Prediction of Synthesizable ... [https://arxiv.org/html/2505.09161v1]
• 64. Optimal pre-train/fine-tune strategies for accurate material ... [https://www.nature.com/articles/s41524-024-01486-1]
• 65. The AI Materials Hype Cycle Hits Its Awkward Teenage ... [https://winsomemarketing.com/ai-in-marketing/the-ai-materials-hype-cycle-hits-its-awkward-teenage-phase]
• 66. Materials Expert-Artificial Intelligence for materials discovery [https://www.nature.com/articles/s43246-025-00928-7]
• 67. A new generative AI approach to predicting chemical ... [https://news.mit.edu/2025/generative-ai-approach-to-predicting-chemical-reactions-0903]