Interpretable Machine Learning for Corrosion Fatigue Life Prediction of Q345C Steel with Scarce Experimental Data

Abstract

Reliable fatigue life prediction for corroded bridge steel remains challenging. Severe surface damage, sparse tests, and heterogeneous literature data obscure the link between corrosion morphology, stress state, and fatigue resistance. In this study, electrochemical accelerated corrosion, three-dimensional laser scanning, tensile testing, axial fatigue testing, and scanning electron microscopy were combined with an interpretable machine learning framework for pre-corroded Q345C steel. A dimensionless feature-aligned support vector regression model was developed by normalizing stress amplitude with material strength, adding Gaussian noise regularization, and integrating multi-source corrosion fatigue data. For 27 experimental validation samples, the model achieved a coefficient of determination of 0.795 and a root mean square error of 0.146, with 26 samples falling within the predefined twofold error band. Shapley additive explanations identified mass loss ratio as the dominant predictor and showed a physically consistent negative effect of normalized stress amplitude on fatigue life. These results suggest that physically informed dimensionless feature alignment can improve small-sample corrosion fatigue prediction while retaining interpretable links to damage mechanisms.