Machine Learning High-Throughput Screening of Double Perovskites for Enhanced Acidic Oxygen Evolution

Abstract

In response to growing demand for cost-effective, highly active, and stable catalysts for the electrocatalytic oxygen evolution reaction (OER), we have developed a comprehensive database for AA’BB’O6-type cubic double perovskites (DP) and implemented a machine learning high-throughput screening (ML-HTS) strategy to discover promising DP for OER. Our approach covers a vast chemical space of approximately 6,500 DP by considering all valid compositional combinations. Additionally, we have developed two machine learning (ML) strategies to predict thermodynamic stability (energy above the hull and Pourbaix decomposition energy) and binding Gibbs free energy, enabling an efficient exploration of the chemical space. Both models, trained with minimal calculations, achieve high accuracies even with unrelaxed structures (mean absolute error (MAE) 0.028 and 0.031 eV/atom for Pourbaix stability and energy above hull, respectively, and MAE 0.124 and 0.129 eV for O* and OH* binding Gibbs free energy, respectively). By utilizing the descriptor-based ML models and graph-based pretrained universal ML potential, we evaluated the stability of approximately 3,500 bulk structures and the activity of around 14,000 surfaces, respectively. This ML-HTS approach led to the successful and efficient identification of 15 novel DPs that are predicted to be more active and stable than established DPs such as LaSrCoFeO6 (LSCF).