Accelerated discovery of OER catalysts in Pnma perovskites via machine learning with minimal DFT structure relaxation

Abstract

The discovery of efficient oxygen evolution reaction (OER) catalysis is essential for advancing sustainable energy technologies. This study presents a machine learning-driven framework to accelerate the identification of alternative OER catalysts, with a focus on multi-metal perovskite oxides composed of Earth-abundant elements. The research integrates traditional experiments with machine learning and theoretical investigations, utilizing descriptors like oxygen p-band center (Op) and metal d-band center (Md). Through high-throughput density functional theory (DFT) calculations and crystal graph convolutional neural networks (CGCNN), the study screens a large compositional space. The key innovation of this work is a framework that predicts the descriptor directly from unrelaxed crystal structures, bypassing the computationally expensive DFT relaxation step and enabling an unprecedented acceleration of the screening process. We predict Op/Md for 149,952 perovskites, highlighting compositions with an Op/Md ratio around 0.48 revealed higher proportions of Ca, Sr, and Ba on the A-site, and Mo, Ni, and Fe on the B-site. This descriptor-based approach offers a computationally efficient and accurate method for screening vast compositional spaces, guiding the rational discovery of promising OER-active perovskite materials.