From descriptors to machine learning interatomic potentials: a review of AI-accelerated electrocatalyst design

Electrocatalysis underpins key energy-conversion reactions, such as carbon dioxide reduction and oxygen evolution/reduction reactions; however, catalyst screening based on density functional theory (DFT) is constrained by rapidly escalating computational costs. In recent years, coupling machine learning (ML) with DFT has opened a promising path around this bottleneck. This review synthesizes recent advances in artificial intelligence (AI)-accelerated electrocatalysis along three pillars: descriptors, ML techniques, and machine learning interatomic potentials (MLIPs). First, we compare the acquisition strategies for three classes of foundational descriptors: intrinsic statistical, electronic structure, and geometric/microenvironmental. Second, we summarize how algorithms such as tree ensembles and kernel methods perform across varying feature dimensionalities and extrapolation regimes, and discuss recent research on customized composite descriptors in the context of these models. Third, we categorize mainstream MLIPs into three families: “general graph-network”, “symmetry-equivariant”, and “extreme-efficiency”. We then compare their trade-offs in accuracy, computational cost, and scope of application. Finally, we highlight limitations of available datasets and propose practical paths forward.