Accurate Long-Term Atmospheric Corrosion Prediction of Weathering Steel via SHapley Additive exPlanations (SHAP) Enhanced Neural Network
Abstract
The growth of the photovoltaic (PV) industry is driving demand for structural materials with superior corrosion resistance, strength, and environmental adaptability. Weathering steels have emerged as promising candidates. However, the long-term atmospheric corrosion performance of weathering steels in complex environments remains insufficiently studied. This paper proposes a data-driven framework to predict corrosion rates of weathering steels based on a neural network model enhanced by SHapley Additive exPlanations (SHAP). The model takes alloy composition, environmental parameters, and exposure time as inputs and outputs the predicted atmospheric corrosion rates. Bayesian optimization is utilized to explore the optimal network configuration and hyperparameter set. SHapley Additive exPlanations is used to identify physically meaningful features, leading to the construction of a SHapley Additive exPlanations -enhanced neural network (ENN). The ENN is evaluated through five-fold cross-validation to assess its robustness and generalization capability, achieving a coefficient of determination (R2) of 0.941. Furthermore, when tested on one-year independent data, the model maintains a low relative error of 2.31%. Moreover, the predicted long-term corrosion trends of WP720 weathering steel under ISO-defined atmospheric conditions align well with established corrosion mechanisms. These results demonstrate the enhanced neural network reliability, high accuracy, and potential for practical deployment in corrosion forecasting and lifespan assessment of weathering steels.