Computational Single-Atom Catalyst Database Empowers the Machine Learning Assisted Design of High-Performance Catalysts

Abstract

The data-driven strategy has emerged as an important approach for the rapid screening of high-performance single-atom catalysts (SACs). However, the lack of a comprehensive SACs database seriously hinders the widespread application of this strategy. Herein, we construct a public SACs database comprising 1197 samples via doping nonmetallic atoms (B, N. O, P, and S) in the coordination environment and regulating 3d metal centers (Ti, V, Cr, Mn, Fe, Co, Ni, Cu, and Zn). Based on density functional theory calculations, the electronic structural properties (i.e., Bader charge and d-band center) and binding energies are obtained. According to the binding energy calculations, 657 stable catalyst configurations are identified. Subsequently, the corresponding adsorption energies for O₂, O, and NO are calculated. Moreover, machine learning (ML) models, specifically extreme gradient boosting regression (XGBR), random forest regression, and support vector regression, are trained to predict the electronic structure and the adsorption energies of O. Among these models, XGBR demonstrates the highest predictive accuracy, with a mean squared error less than 0.35. We successfully integrate ML models based on this SACs database and catalytic volcano model. Through this framework, the catalytic activities of 1261 4d SACs in the oxidation of NO and Hg⁰ are quickly predicted. Rh₁B₄ and Rh₁C₂S₂ are identified as potential catalysts for the oxidation of NO and Hg⁰, with the respective energy barriers of 1.01 and 2.59 eV for Rh₁B₄, and 1.03 and 2.61 eV for Rh₁C₂S₂. These values are significantly lower than those of previously reported SACs. We anticipate that this public SACs database and ML-based activity prediction framework can provide new pathways for the rapid screening of highly active SACs for various catalytic reactions.