Predictive Modelling of Steel Corrosion: A Comparative Study of Statistical and Machine Learning Methods with a Focus on Copper Behavior

Abstract

Accurately predicting corrosion rates in low-alloy steels is a significant challengein materials engineering due to the intricate and nonlinear interaction between environmental exposure conditions and alloying elements. Conventional statistical corrosion models are by and large based on linear assumptions, and thereby most of them fail to capture such interactions, especially the dual and threshold-controlled behavior of Cu in forming protective rust layers. This study presents a comprehensive comparative analysis of statistical and machine learning approaches for the prediction of atmospheric corrosion rates of low-alloy steels. Several predictive models were developed and validated for this task, namely, Decision Trees, Support Vector Machines with several kernel functions, and the gradient-boosting algorithm XGBoost, by using experimentally derived environmental and compositional datasets. Model performance is assessed based on the coefficient of determination (R²), root-mean-square error (RMSE), and mean absolute error (MAE).The results demonstrate that non-linear models outperform conventional linear models of non-linear models in comparison to the conventional linear models. Among various models, the best predictive model was XGBoost with an R² value of about 0.88, validating the best ability to exploit threshold values and interactive coefficients. The conclusion is drawn from the result that copper is only effective in imparting corrosion resistance in a critical range, beyond which the protective property becomes saturated or decreases, an imperative that cannot be accounted for in the linear regression models. The obtained results strongly suggest that the corrosion process in low-alloy steels is governed by nonlinear and threshold-dependent behavior. Therefore, the paper establishes the application of gradient - boost-based machine learning algorithms as a strong and authentic method for predicting the development of corrosion models in low-alloy steels. Furthermore, the comparative assessment also emphasizes the importance of kernel-based learning approaches, such as the radial basis function kernel-based SVM, where the models also forecast well with slightly lower accuracy compared to the gradient-booster models. Additionally, the better performance of the non-linear models clearly depicts the model flexibility to cope with the Abrupt transitions of the corrosion process influenced by varying concentrations of alloys. From the scientific aspect, the obtained results demonstrating the threshold effects of the copper concentrations clearly support the formation and saturation of inner Cu-enriched rust layers. Therefore, this analysis attempt to connect the machine learning algorithm results with the scientific concepts of the corrosion process and hence establish the importance of machine learning models as a supplementary aspect to the other scientific methods used in the fields of materials science.