[17a-S2_204-1~13] 23.1 Joint Session N "Informatics"

9:45 AM - 10:00 AM JST(12:45 AM - 1:00 AM UTC)

[17a-S2_204-4] A High-Efficiency Materials Design Method Using Pretrained Models and Their Gradients

〇Akihiro Fujii¹, Yoshitaka Ushiku², Koji Shimizu³, Anh Khoa Augustin Lu^4,1, Satoshi Watanabe¹ (1.UTokyo, 2.OMRON SINIC X Corp., 3.AIST, 4.NIMS)

Keywords:

materials design,generative model,machine learning

Designing novel materials with useful properties is a critical challenge in materials science, and in recent years deep generative models trained on large-scale datasets have become the dominant approach. In this study, we demonstrate that flexible materials design can be achieved even with limited data by directly optimizing material representations using gradients of pretrained property prediction models, without relying on generative models. Our approach can be applied without retraining existing models and achieves a higher success rate in proposing novel and stable structures than a leading generative model trained on much larger datasets.