Fine-structure sensitive deep learning framework for predicting catalytic properties with high precision

Abstract

The fine structure of a surface considerably affects its catalytic performance in structurally sensitive reactions. High-throughput (HT) screening and machine learning (ML) are considered efficient for exploring the hidden rules of impacts. However, no protocol for constructing an interpretable ML framework sensitive to fine structures has been reported thus far. Herein, we developed a data augmented convolutional neural network (CNN)-based ML framework called “global + local” convolutional neural network (GLCNN), which combines “global + local” features. This framework captures original fine structures without the use of complicated encoding methods by transforming the catalytic surfaces and adsorption sites into two-dimensional grids and one-dimensional descriptors, respectively. The GLCNN framework accurately predicted and distinguished the adsorption energies of OH on a set of analogous carbon-based transition-metal single-atom catalysts with a mean absolute error of less than 0.1 eV. Moreover, this model yields the best results among popular models trained on large datasets so far. Unlike conventional CNN and descriptor-based models with one-sided feature extraction, this fine-structure-sensitive ML framework can extract key factors that affect the catalytic performance from both geometric and chemical/electronic features, such as symmetry and coordination elements, through unbiased interpretable analysis. This framework provides a feasible solution for the high-precision HT screening of heterogeneous catalysts with a broad physical and chemical space.