Explainable Synthesizability Prediction of Inorganic Crystal Polymorphs Using Large Language Models

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Predicting the synthesizability of hypothetical crystal structures is crucial for efficient materials discovery.
• Current methods often rely on complex, specialized models that may lack interpretability.

Purpose of the Study:

• To evaluate machine learning, specifically large language models (LLMs), for predicting crystal structure synthesizability.
• To develop an interpretable AI workflow for materials design guidance.

Main Methods:

• Fine-tuning LLMs on text descriptions of crystal structures.
• Employing positive-unlabeled learning with text-embedding representations.
• Developing an LLM-based workflow for explanation generation and rule extraction.

Main Results:

• LLMs achieve comparable performance to graph neural networks for synthesizability prediction.
• Positive-unlabeled learning on text embeddings improves prediction quality.
• The LLM workflow successfully generates human-readable explanations and extracts physical rules.

Conclusions:

• AI, particularly LLMs, can effectively predict material synthesizability and provide interpretable insights.
• This approach offers a powerful tool to guide chemists in optimizing hypothetical structures for real-world applications.
• The developed workflow enhances the feasibility of materials design by offering actionable guidance.