Data-Driven Corrosion Science: Artificial Intelligence, Computational Chemistry and Predictive Electrochemistry

Machine learning has become a standard analytical instrument in corrosion science, yet the transition from predictive demonstration to dependable engineering use remains incomplete. This review critically examines the four properties on which that transition depends: the quality and provenance of corrosion data, the interpretability of fitted models, the quantification and calibration of predictive uncertainty, and the consistency of learned relationships with established corrosion mechanisms. Literature was identified through structured searching of open scholarly indexes, supplemented by backward and forward citation tracking and by checks for corrections and retractions, with all bibliographic details and digital object identifiers verified against registration records. The synthesis indicates that trustworthiness in this field is frequently framed as a problem of model transparency, whereas the binding constraint lies earlier, in datasets assembled from heterogeneous exposure programmes, inconsistent electrochemical measurement practice and inconsistently defined degradation outcomes. Attribution-based explanation, overwhelmingly Shapley additive explanation, now accompanies most corrosion models, but its outputs are conditional on the fitted model and on strongly collinear environmental descriptors, so recovered feature rankings function as hypotheses rather than mechanistic findings. Uncertainty treatment is the least developed dimension: probabilistic outputs are reported in a minority of studies, calibration is rarely assessed, and performance under distribution shift, which is the operative condition for most asset applications, is seldom tested. Mechanistic consistency is generally assessed after fitting by comparing attributions with expert expectation, rather than imposed as a constraint, even though corrosion offers unusually firm kinetic and thermodynamic structure that could be encoded directly. Synthetic data augmentation, now common in small-sample corrosion studies, introduces a further circularity that current evaluation designs do not resolve. Progress requires provenance-complete and openly documented exposure datasets, explanation claims evaluated against mechanistic reference cases, calibrated and distribution-shift-aware uncertainty reporting, and architectures in which corrosion kinetics constrain rather than merely inform the learned function.