Feature Sensitivity and Robustness in Corrosion Rate Forecasting: A Comparative Study of Deterministic and Probabilistic Machine Learning Models

Accurate corrosion rate prediction is critical for ensuring industrial infrastructure safety and reducing economic losses, yet existing models often fail to balance precision, uncertainty quantification, and adaptability to dynamic scenarios. This study systematically compared the Extra Trees Regressor (ETR) and an Improved Bayesian Neural Network (ImprovedBNN) across static performance metrics (MSE, MAE, coverage), feature sensitivity to ±10% fluctuations, and statistical robustness via paired t-tests. The results show that ETR outperformed in point prediction (MSE=0.0051) while ImprovedBNN achieved perfect uncertainty coverage (1.000) with enhanced sensitivity to critical features, with both models demonstrating distinct engineering applicability. These findings establish a framework for model selection in corrosion management, reconciling deterministic precision and probabilistic rigor to inform practical decision-making.