[15p-S2_204-1~18] 23.1 Joint Session N "Informatics"

5:15 PM - 5:30 PM JST(8:15 AM - 8:30 AM UTC)Presentation by Applicant for JSAP Young Scientists Presentation Award

[15p-S2_204-14] Effects of Structural Representations and Vocabulary Design on Crystal Property Prediction with Large Language Models

〇Shuichiro Ozawa¹, Izumi Takahara¹, Teruyasu Mizoguchi² (1.Sch. of Eng., UTokyo, 2.IIS, UTokyo)

Keywords:

Large Language model,Crystal property prediction,Crystal structure description

We investigate the prediction of formation energies of crystalline materials using large language models (LLMs), focusing on (1) natural-language representations of crystal structures and (2) the effect of domain-specific tokenization that treats materials-science terminology as single tokens on prediction accuracy. In particular, we show that fine-tuning is effective when using detailed structure descriptions that include not only composition and space group but also atomic coordination environments.