After pretraining a large language model on synthesizability information for inorganic materials, materials predicted to be difficult to synthesize are redesigned into synthesizable forms. This process implements a human-like "learn and redesign" strategy in which knowledge is acquired and then applied. Credit: Journal of the American Chemical Society
A research team led by Prof. Yousung Jung of the Department of Chemical and Biological Engineering at Seoul National University (SNU) has developed an innovative AI-based technology that uses large language models (LLMs) to redesign new materials that were previously difficult to synthesize into forms that are experimentally feasible.
From prediction to practical redesign
Rather than stopping at the conventional task of predicting a material's synthesizability, the study goes a step further by presenting a practical solution for redesigning materials deemed difficult to synthesize. The technology is expected to significantly accelerate the development of novel advanced materials, including next-generation semiconductor materials and high-efficiency battery materials.
The research was conducted with Jaehwan Choi, an integrated MS–Ph.D. student at SNU, and Dr. Seongmin Kim, a postdoctoral researcher, serving as co–first authors. The results were published on October 6, 2025, in the Journal of the American Chemical Society.
The bottleneck in materials synthesis
Advances in computational chemistry and artificial intelligence have enabled researchers to identify large numbers of theoretically promising material candidates. However, the challenge of actually synthesizing these materials in the laboratory has remained a major bottleneck. While previous research has focused on predicting synthesizability, it has not addressed how to transform materials deemed difficult to synthesize into synthetically feasible structures.
To address this limitation, the research team developed a new LLM-based framework called "SynCry." This model represents the crystal structures of new materials as invertible textual descriptions and learns, through iterative fine-tuning, how to transform structures predicted to be difficult to synthesize into ones that are experimentally feasible.
Redesign results and validation
The results showed that SynCry began with 514 successful structure transformations and, through iterative fine-tuning, ultimately succeeded in redesigning 3,395 structures into synthesizable forms. Particularly noteworthy is that 34 of the top 100 redesigned structures, despite not being present in the training dataset, matched materials that have been experimentally synthesized and reported in the scientific literature. This demonstrates that SynCry goes beyond mimicking training data and is capable of generating genuinely new, synthesizable material structures.
This redesign capability demonstrates that LLMs can move beyond simple prediction to become practical tools for materials design through a "learn-and-regenerate" strategy. The approach has the potential to dramatically shorten the development timeline for advanced materials and to recover numerous candidate materials that were previously discarded due to synthesis difficulties.
Prof. Yousung Jung stated, "This study is the first to demonstrate that AI can directly redesign new materials starting from structures that are difficult to synthesize. We plan to expand this work to a wider range of material systems and larger datasets, ultimately developing a practical tool for discovering new materials."
Jaehwan Choi, an integrated MS–Ph.D. student, commented, "This research began with a simple question: could we bring back virtual materials that had been discarded because they were considered too difficult to synthesize? Moving forward, I aim to develop general-purpose AI agents, including language models, to further accelerate materials discovery."
Dr. Seongmin Kim, who has long conducted research on LLM-based synthesizability prediction, added, "This achievement is an important example showing that AI can play a creative design role in materials science."
Jaehwan Choi plans to pursue research on developing general-purpose AI agents, including large language models, to automate the identification of synthesis mechanisms and optimal synthesis pathways for inorganic materials. Dr. Seongmin Kim, currently working at the SNU Institute of Advanced Chemical Engineering, plans to continue follow-up research that integrates machine learning and materials science to explore paradigm shifts in new materials development.
