The research team of Kumada-Sato-Fujii-Umemoto Laboratory, including Hajime Shimakawa (D3), Associate Prof. Masahiro Sato, and Prof. Akiko Kumada, has developed a quantum mechanics-assisted machine learning model to achieve an extrapolative property prediction for unprecedented materials. Their proposed model enables extrapolative property predictions solely from small experimental data, which overturns the common understanding of AI relying on big data. The model transforms molecular structures into physical quantities obtained through quantum chemical calculations and categorical features and extracts complex relationships between quantum chemical information and material properties. The study conducted a large-scale benchmark test using 12 types of experimental datasets of numerous organic compounds to evaluate the extrapolation performance of various models. The results demonstrate that the proposed model excels in both interpolation and extrapolation performance for property prediction, establishing its suitability as an optimal model for material exploration. This research addresses the challenge of extrapolative material exploration in Materials Informatics and is expected to contribute to the creation of unprecedented materials surpassing existing ones. The findings were published online in the Nature series scientific journal, ‘npj Computational Materials,’ on January 10th.
(a) Extrapolability of property prediction necessary for materials exploration
(b) Difference between this study and previous studies focusing on data size and extrapolability
<Article>
Hajime Shimakawa, Akiko Kumada and Masahiro Sato, “Extrapolative prediction of small-data molecular property using quantum mechanics-assisted machine learning”, npj Computational Materials, 10, 11 (2024).
https://doi.org/10.1038/s41524-023-01194-2
Press release by Faculty of Engineering: https://www.t.u-tokyo.ac.jp/en/press/pr2024-01-30-002