Sim2Real Transfer for Catalyst Activity Prediction

Abstract

Sim2real transfer, knowledge transfer from computational data to experimental data, has received increasing attention as a promising solution to small data problems in materials. We proposed a sim2real transfer method that significantly enhances catalyst activity predictions by harnessing the knowledge of catalyst chemistry. The proposed method transforms the feature space of source computational data into that of target experimental data, and then solves the problem as a homogeneous transfer learning. Through the demonstration, we confirmed that transfer learning model exhibits positive transfer on accuracy and robustness. Notably, significantly high accuracy was achieved despite using a few (less than 10) target data, whose accuracy is compatible with a full scratch model with more than 70 target data. This result indicates that the proposed method leverages the prediction performance with few target data, which helps saving the number of trials in real laboratories.