The rational design of high-performance graphene-based single-atom electrocatalysts for the ORR using machine learning.

Abstract

In this work, high-performance two-dimensional (2D) graphene-based single-atom electrocatalysts (ZZ/ZA-MNxCy) for the oxygen reduction reaction (ORR) were screened out using machine learning (ML). A model was built for the fast prediction of electrocatalysts and two descriptors valence electron correction (VEc) and degree of construction differences (DC) were proposed to improve the accuracy of the model prediction. Two evaluation criteria, high-performance catalyst retention rate rR and high-performance catalyst occupancy rate rO, were proposed to evaluate the accuracy of ML models in high-performance catalyst screening. The addition of VEc and DC in the model could change the mean absolute error (MAEtest) of the test set, the coefficient of determination (R2test) of the test set, rO, and rR from 0.334 V, 0.683, 0.222, and 0.360 to 0.271 V, 0.774, 0.421, and 0.671, respectively. The partially screened potential high-performance ORR electrocatalysts such as ZZ-CoN4 and ZZ-CoN3C1 were also further investigated using a Density Functional Theory (DFT) method, which confirmed the accuracy of the ML model (MAE = 0.157 V, R2 = 0.821).