Interpretable machine learning reveals a morphological regime shift governing pitting resistance in steam-coated boehmite films

Steam coating forms protective boehmite films on aluminum alloys; however, identifying the primary factors governing corrosion resistance remains challenging due to strong descriptor covariance. We established a physics-informed interpretable machine-learning framework to predict pitting potential ( E pit ) from four physical descriptors: film thickness (FT), surface morphology (SQ), crystallite size (CS), and substrate dislocation density index (FS) using Random Forest ( R ² = 0.67). SHAP analysis identified SQ and FT as dominate descriptors, while ALE analysis showed that SQ’s role shifts from marking protective coverage in early-growth films to defect-mediated degradation in mature films. Two-dimensional ALE analysis indicated that corrosion resistance arose from region-dependent synergistic interactions among film geometry, crystallinity, and substrate defect structure, with the SQ–CS and FT–SQ pairs showing sign-reversing interaction landscapes rather than a single uniform synergy. Monte Carlo simulations confirmed that measurement-derived uncertainty (~0.080 V) accounts for only a minor fraction of the total model error, indicating residual errors stem mainly from unobserved physical heterogeneities. These findings demonstrate how interpretable machine learning can decouple intertwined physical descriptors and propose candidates for degradation-governing mechanisms in complex process–structure–property systems.