AI models transform material designs by predicting and explaining synthetic properties

The researchers successfully developed techniques to use large-scale language models (LLMs) to predict the synthesis potential of new materials and interpret the basis of such predictions. The team was led by Professor Yuseun Jung of Seoul National University and was led in collaboration with Fordham University in the United States.

The results of this study are expected to contribute to the new material design process by either excluded candidates that are less likely to be synthesized in advance, or by optimizing previously challenged materials to a more feasible shape.

This study was published in two chemical journals: Postdoctoral researcher Sungmin Kim and first author, in the American Chemistry Society Journal on July 11, 2024 and in the Angewandte Chemie International Edition on February 13, 2025.

Accurately assessing the feasibility of synthesizing materials is important when developing new materials. If material designs do not properly consider synthesisability, they can lead to unnecessary experiments on unverified virtual structures, resulting in inefficient use of research resources and time. This underscores the need for accurate synthesisability prediction techniques.

However, existing prediction methods were limited to assessing the thermodynamic stability of materials, resulting in a significant discrepancy between prediction and the success rate of actual experimental synthesis. Although machine learning models have been developed to address this issue, they primarily focus on classification without explaining the rationale behind their predictions, thus lacking explanability and reliability.

To overcome these challenges, Professor Jung’s research team discovered that LLM can not only accurately predict the synthesis of inorganic crystal polymorphisms, but also ensure explanatory potential.

The researchers first fine-tuned the generic LLM using an inorganic crystal material data set in a text-based format. This model was then trained to classify the synthetic nature of a particular virtual material, predict the precursor compounds required for synthesis, and identify and interpret key factors affecting synthesisability. As a result, LLM achieved a higher level of prediction accuracy than existing bespoke machine learning models.

Furthermore, the team found that LLM could go beyond mere predictions and provide an interpretable explanation of why certain material is synthesisable. This breakthrough opens the door to analyzing why it is difficult to synthesize a particular hypothetical crystal structure and why it is difficult to identify factors that hinder synthesis. Furthermore, this study uncovered complex previously unknown correlations and factors that influence the feasibility of material synthesis.

This groundbreaking technology for predicting and explaining synthesis potential is expected to make a significant contribution to the domestic advanced materials industry and to increase the competitiveness of the semiconductor and secondary battery industries. While traditional new methods of discovering materials include numerous trial and error experiments, LLM-based prediction techniques can accelerate material design and reduce development time.

Furthermore, this work can be applied to the design of semiconductor devices and highly efficient battery materials, and is expected to help maintain Korea’s technical leadership in advanced materials and ensure an early market advantage. When commercialized, this work could serve as an important tool for research institutions and companies to quickly identify new materials and assess the potential for mass production.

Professor Yuseun Jung said, “This study is important because LLM can not only predict the possibility of accurate new materials synthesis, but also interpret the reasons behind those predictions and uncover the underlying chemical principles.

“As LLM-based technology continues to evolve, it is expected to provide a more efficient and intuitive orientation for new material designs.”

Kim, a postdoctoral researcher at the Institute of Chemical Processes at Seoul National University, will conduct follow-up research that integrates machine learning and materials science to explore a paradigm shift in new material development.