Atmospheric Corrosion Prediction of Low-Alloy Carbon Steel Based on Dual-Correction Dynamic Ensemble Framework

Abstract

Accurate forecasting of atmospheric corrosion for low-alloy carbon steel is critical for lifecycle management of infrastructure, yet data-driven models often exhibit systematic bias under highly variable environments. To address the limitations of atmospheric corrosion prediction—particularly systematic prediction bias and residual skewness commonly observed in gradient-boosting-based learners—this study proposes a Dual-Correction Dynamic Ensemble (DCDE) framework. DCDE integrates XGBoost and CatBoost in a two-layer architecture, where a primary prediction layer captures baseline nonlinear degradation behavior and a residual correction mechanism compensates structured errors. A unified Bayesian optimization protocol is adopted to jointly configure model structures, correction depth, and fusion parameters, enabling synergistic interaction between the boosting components. Experimental validation on industrial atmospheric corrosion datasets using 5-fold cross-validation demonstrates the effectiveness of DCDE: the proposed method achieves RMSE = 0.2988 ± 0.0473 and Pearson R = 0.9627 ± 0.0138, and reduces the mean relative error by approximately 10.6% compared with the baseline XGBoost model. The main contributions include (i) synergistic integration of heterogeneous boosting architectures, (ii) a dual-correction ensemble design that improves error compensation and stability, and (iii) a joint Bayesian optimization scheme that enhances predictive accuracy and interpretability for corrosion-driven lifespan prediction in complex degradation environments.