Mechanistic insights into supercritical CO2 corrosion via interpretable Kolmogorov-Arnold networks

Predicting steel corrosion in supercritical CO 2 (ScCO 2 ) pipelines is important for carbon capture, utilization, and storage (CCUS) systems but challenging, because many coupled variables are involved. Machine learning can reach reasonable accuracy, yet most models are black boxes offering little mechanistic insight. This study applies the Kolmogorov–Arnold Network (KAN), whose learnable edge functions can be read directly as curves, to predict corrosion rate and pitting of steels in ScCO 2 . Using 385 samples from 32 published studies under a leak-free cross-validation protocol, the regression model reached a mean test R 2 of 0.82, comparable to tuned tree ensembles, so the KAN is competitive rather than more accurate; the classifier reached 89.2% accuracy and 0.948 AUC. Because prediction and interpretation share one model, its partial-dependence curves were read directly: chromium showed saturating protection, temperature an Arrhenius-type increase, immersion time a power-law decay, and pressure a non-monotonic trend, all broadly consistent with corrosion chemistry, while the increasing oxygen effect reflects enhanced mass transport under dynamic conditions. Unlike post-hoc attributions, which depend on the choice of method and background, the KAN curves are smooth and uniquely defined, making it a practical tool for studying ScCO 2 corrosion mechanisms