[17p-S2_204-1~18] 23.1 Joint Session N "Informatics"

5:00 PM - 5:15 PM JST(8:00 AM - 8:15 AM UTC)Presentation by Applicant for JSAP Young Scientists Presentation Award

[17p-S2_204-13] Inverse design of surfaces using a data-driven approach

〇(M1)Kazuya Miyamoto¹, Ibuki Okuda¹, Teruyasu Mizoguchi¹ (1.Tokyo Univ. IIS)

Keywords:

material informatics,data-driven analysis,machine learning

In this study, we propose an inverse design methodology for lattice defects based on deep generative diffusion models . By integrating high-quality datasets derived from machine learning interatomic potentials with our "Modified MatterGen"—enhanced with an atom-site labeling function—we enable the direct generation of optimal defect configurations based on target physical properties . This framework facilitates the autonomous creation of structures even in out-of-distribution regions while strictly maintaining geometric consistency. We report on these achievements, which accelerate defect engineering while dramatically reducing the computational costs inherent in traditional forward-search approaches, alongside case studies involving various crystal surfaces and interfaces .