Evidential Crystal-Graph Convolutional Neural Networks for Efficient Global Optimization of Electrocatalysts

Trial-and-error experimental approaches to catalyst discovery are time consuming and expensive. Computational screening of catalysts using quantum mechanical methods such as density functional theory (DFT) modeling are highly useful, but these approaches are still limited because of the high computational cost. Machine learning (ML) using deep neural networks has emerged as a powerful complement to high-throughput screening, enabling prediction of catalysts at much faster speeds compared to using only experimental methods or DFT modeling. However, the full utility of ML models to explore large catalyst spaces can only be obtained if it is straightforward to identify when model predictions are very uncertain or accurate, Fig. 1. Established uncertainty quantification approaches for neural networks are typically costly to obtain and have limitations in evaluating prediction errors for catalyst space exploration. To enable broader use of ML for catalyst discovery, new uncertainty quantification approaches must be developed. We will implement evidential regression with crystal-graph convolutional neural networks (CGCNN) to enable accurate prediction of model uncertainty and accelerate electrocatalyst optimization for energy applications. We will use evidential regression CGCNN within an optimization framework to discover electrocatalysts for sustainable fuel.