From prediction to materials design: machine learning in electrocatalytic water splitting

Machine learning (ML) is emerging as a powerful strategy to accelerate electrocatalyst discovery for water splitting, yet its impact is still limited by data quality, thermodynamic bias, and weak coupling to realistic experiments. This perspective is a survey of recent ML-guided efforts on both oxygen and hydrogen evolution reactions, with emphasis placed on compositionally complex platforms such as layered double hydroxides, metal-organic frameworks, high-entropy alloys and related multimetal systems. It highlights how supervised models, ML-density functional theory (DFT) workflows and interpretable surrogates are being used for rapid composition screening, dopant optimization and non-linear descriptor discovery beyond simple volcano-type relations. At the same time, this article identifies structural bottlenecks including small, compositionally clustered datasets, overreliance on static adsorption energies, underrepresentation of kinetics and stability, and a persistent gap between computational predictions and performance at industrially relevant current densities. Finally, this Perspective outlines opportunities for integrating active learning, operando-informed labels, physics-based and microkinetic models, and explainable artificial intelligence (SHapley Additive exPlanations, symbolic regression, and counterfactual design) into closed-loop, synthesis-aware workflows, positioning ML as a mechanistically meaningful design tool for robust, earth-abundant water-splitting electrocatalysts.