Presentation Information
[25p-61C-12]Polymer exploration for electrical equipment using self-supervised deep learning model
〇Chihiro Tateyama1, Gen Komiya1, Hiroyasu Tarui1 (1.Toshiba Infrastructure Systems & Solutions Corp.)
Keywords:
materials informatics,self-supervised deep learning,material exploration
Compared with metals and inorganic materials, resin materials have complex structures and it is difficult to collect a large amount of experimental data of resin required to conduct materials informatics. There is a room for inprovement of existing prediction models to achieve high prediction accuracy because machine learning models require appropriate feature selection and simple supervised deep learning models require a large amount of data. In order to enable highly accurate property prediction even with a small amount of experimental data, we constructed the prediction model that employs a self-supervised deep learning model that combines unsupervised pretraining which learns molecular structure rules, and fine tuning which constructs a property prediction model from small experimental data. We also investigated the material exploration scheme and interpretability based on the prediction model.