Data-driven corrosion assessment of magnesium alloys: A multi-level graph attention network for quantitative hydrogen-evolution prediction

Magnesium (Mg) alloys are highly valued in aerospace, biomedical and other fields due to their high specific strength. However, non-uniform corrosion failure during service remains a core challenge that restricts their engineering applications. Traditional corrosion kinetics models fail to accurately elucidate the cross-scale synergy mechanism between microstructure and macroscopic corrosion behavior. In this study, based on 13 kinds of Mg alloys, 20 sets of 100-h hydrogen evolution curves, and characterization data from scanning electron microscopy (SEM) and electron backscatter diffraction (EBSD) information, a multi-level corrosion kinetics database was constructed, covering physicochemical parameters, micro-grain topological structures and second phase features, as well as macroscopic statistical characteristics and temporal dimension. Through machine learning algorithms, key corrosion driving factors were identified, and a multi-level graph attention network modeling framework was proposed, where the grains and grain boundaries were constructed as a graph structure, and the hierarchical interaction modeling between microstructure and corrosion kinetics was realized by combining the attention mechanism. The model has been validated in a new Mg alloy dataset for its predictive capability across compositional systems. This work provides a new computational paradigm and significantly enhances the predictability and efficiency of corrosion-resistant Mg alloy design.