[8a-N302-1~11] 18.1 Materials Informatics

9:45 AM - 10:00 AM JST(12:45 AM - 1:00 AM UTC)Presentation by Applicant for JSAP Young Scientists Presentation Award

[8a-N302-4] Estimating Reliable Feature Importance by Integrating Generative AI Prior Knowledge and Bayesian Model Averaging for Materials Discovery

〇Yuki Namiuchi¹, Yuka Kitamura², Kan Hatakeyama³, Yuya Oaki², Yasuhiko Igarashi¹ (1.Tsukuba Univ., 2.Keio Univ., 3.Tokyo Univ.)

Keywords:

sparse modeling,large language model,LLM posterior inclusion probability

Materials exploration requires feature selection that remains effective even with small datasets. In this study, we propose a method that selects features based on posterior probabilities, integrating feature prior probabilities generated by an LLM with the data contribution obtained through Bayesian model averaging. Validation on nanosheet material yield showed that priors based on selection frequency captured differences between features better than those based on likelihood, and by updating the posterior probabilities during exploration, the method achieved materials exploration performance surpassing both the use of all features and static selection.