AI-Driven Electrocatalyst Discovery: Integrating Machine Learning, Density Functional Theory, and High-Throughput Screening for Sustainable Energy Conversion

Abstract

Efficient electrochemical processes are essential for a sustainable, low-carbon energy economy, involving hydrogen production, oxygen evolution/reduction, and carbon dioxide valorisation, all of which require high-performing electrocatalysts. Traditional catalyst development methods, which involve sequential trial-and-error synthesis and density functional theory (DFT) calculations, cannot keep pace with the combinatorial complexity of multi-metallic, single-atom, and high-entropy alloy catalysts. This review summarizes recent advances in AI-driven electrocatalyst discovery in the past five years (2020–2025) with a focus on the three pillars that have enabled this development: (1) machine learning (ML) algorithms and interpretable/explainable frameworks that extract descriptors of catalytic data that are physically meaningful; (2) the ability to couple ML with DFT using surrogate models, graph neural network interatomic potentials, and large benchmark datasets that reduce computational costs by orders of magnitude at close to DFT accuracy; and (3) high throughput screening pipelines for virtual and experimental screening, involving active learning and Bayesian optimisation, that have been applied to a diverse range of electrocatalytic reactions, including the hydrogen evolution reaction, the oxygen evolution/reduction reaction, and the carbon dioxide reduction reaction. Single-atom catalysts, perovskite oxides, and high-entropy alloys are discussed as representative examples that have helped discover new, high-performance candidates and reduced the need for one to two orders of magnitude of DFT calculations. The review also covers emerging autonomous, closed-loop ‘selfdriving’ laboratories, where robotised synthesis and characterisation are coupled with ML-driven decision-making. The review critically discusses persistent challenges in data scarcity, model interpretability and transferability across chemical spaces, and experiment-theory reproducibility, and envisions future directions such as foundation models for atomistic simulation and multimodal data fusion toward fully autonomous electrocatalyst discovery for sustainable energy conversion.