Dynamic Machine Learning for the Oxygen Evolution Reaction: From Static Catalyst Screening to Working-State Trajectory Learning

Efficient water electrolysis is limited not only by the sluggish kinetics of the oxygen evolution reaction (OER), but also by the dynamic reconstruction and degradation of catalysts under operating bias. Machine learning (ML) has accelerated OER catalyst screening by linking precursor composition, structural descriptors and adsorption energetics to endpoint activity, yet static endpoint models cannot determine when a working interface forms, how reaction pathways redistribute, or where stability costs accumulate during operation. This Review does not treat catalyst reconstruction or ML-assisted operando analysis as new concepts; instead, it formalizes their intersection as a supervised trajectory-learning problem. We define dynamic OER ML operationally as learning in which ordered operating history enters the prediction and the target contains at least one time-resolved state, transition event, pathway redistribution, degradation event or transfer trajectory. The field is organized around three learning targets: active-interface formation from precursors, coupled evolution of OER mechanisms and degradation routes, and transfer of material behavior across pH, electrolyte, potential program, transport and device boundaries. Static ML is repositioned as a source of precursor priors, whereas dynamic ML must learn state transitions, rates, latent working states, uncertainty and protocol-dependent transfer from time-resolved evidence. We further provide operational routes from raw operando signals to quantitative labels, target-specific evaluation metrics, baseline comparisons and leakage-safe validation rules. By shifting OER learning from material-endpoint ranking to material-condition-trajectory prediction, dynamic ML can support the prediction, preservation and control of real catalyst working states.