Presentation Information

[9a-S301-2]Sim2Real transfer learning with chemistry-informed domain transformation

〇Yuta Yahagi1,2, Kiichi Obuchi1,2, Fumihiko Kosaka2, Kota Matsui3 (1.NEC, 2.AIST, 3.Kyoto Univ.)

Keywords:

machine learning,first principles calculation,catalyst

We propose a transfer learning approach that leverages large first-principles calculation datasets. It can improve predictive accuracy even when only a small amount of experimental data is available. The key innovation is a chemistry-informed domain transformation. This enables knowledge transfer between computational and experimental domains by bridging their domain heterogeneity. We carried out catalyst activity prediction to validate this method, resulting an order-of-magnitude improvement over models without transfer learning. These results clearly demonstrate positive transfer.