Data-Driven Prediction and Interpretability Analysis of Corrosion Rate for Low-Alloy Steel

Abstract

This study focuses on low-alloy steels for offshore equipment in islands and coastal regions, which are exposed to aggressive marine atmospheres. A corrosion rate prediction method integrating data augmentation, feature selection, and interpretable machine learning is proposed. The conditional tabular generative adversarial network (CTGAN) is employed to generate synthetic samples on the training subset with rigorous quality validation. The predictive performance of six machine learning models is compared, and four feature selection strategies are applied to construct optimized datasets. Meanwhile, the Shapley additive explanations (SHAP) method is introduced to perform both global and local interpretability analysis on the optimal model. Experimental results demonstrate that the extreme gradient boosting (XGBoost) model, combined with the feature subset selected by gradient boosting decision tree (GBDT)-based importance evaluation, achieves the best performance, with a coefficient of determination (R2) of 0.9518 and a mean absolute error (MAE) of 2.1697 μm·a−1. The model’s generalization to unseen exposure stations was evaluated via leave-one-station-out validation. SHAP analysis indicates that exposure time, SO2 deposition rate, and mass fractions of Cr and P are the key factors, with influence directions consistent with marine atmospheric corrosion theory. This approach provides a reliable corrosion-prediction tool for low-alloy steels under dataset-similar marine-atmospheric conditions with limited samples.