Machine Learning-Assisted Active Center Exploration in Atomically Thin MoS_xTe_2-x Electrocatalysts for Efficient Hydrogen Evolution.

Abstract

Modulating the local configurations is widely considered an efficient strategy to promote the catalytic performance of 2D molybdenum disulfide (MoS₂) for hydrogen evolution reaction (HER). Although transmission electron microscopy prevails as a central tool to visualize catalysts at atomic resolution, there still lacks a rapid and accurate approach to finding the active centers in the micrographs containing abundant structural information. Herein, a defective MoS_xTe_2-x alloy catalyst is created through low-temperature sulfurization of 1T'-MoTe₂ (S-MoTe₂), whose atomic structure is automatically explored using an unsupervised machine learning (ML) framework based on the Zernike feature and uniform manifold approximation and projection (UMAP)-assisted clustering, enabling the discovery of a novel defect configuration referred antisite Te adatom (Te_ads-Mo). Density functional theory (DFT) calculations reveal a synergistic enhancement in both the hydrogen adsorption capability and electronic conductivity of these antisite defects, which is experimentally verified by the half-reduced overpotential and Tafel slope of S-MoTe₂ alloy compared to its counterparts without Te_ads-Mo. This work provides an intelligent approach to facilitate active center exploration in micrographs and achieves a closed-loop verification for the ML-assisted defect discovery via theoretical calculations and electrochemical experiments, displaying how ML and researchers seamlessly cooperate in a scientific workflow for advanced catalyst development.