Machine learning-enabled corrosion prediction and feature importance analysis of Al6061 stir-cast metal matrix composites using SHAP interpretability

Abstract This work investigates the application of machine learning techniques for predicting the corrosion properties of Al6061 stir-cast metal matrix composites, including Corrosion Rate (CR), corrosion potential (Ecorr), and corrosion current density (Icorr). Machine learning regression models such as SVR, XGBoost, RF, BNN, ANN, and CatBoost are used. The efficiency of each model in predicting the outcome variables is determined by utilizing statistical measures of prediction accuracy. It is found that among all the models, the best prediction results are obtained by the XGBoost with an R2 value of 0.9931 for CR, 0.9896 for Ecorr, and 0.9914 for Icorr. The SHAP analysis is employed to find the effect of process variables on corrosion behavior to improve the interpretability of the machine learning outputs. From the SHAP analysis, it is evident that the most important factors affecting the corrosion behavior are the casting temperature and reinforcement weight percentage, followed by particle size. The stirring speed and stirring time have comparatively lower effects on the corrosion behavior and Ecorr, Icorr, and corrosion rate increased with an increase in casting temperature and reinforcement percentage. Overall, the present research provides a systematic analysis of several machine learning techniques used in predicting corrosion properties of Al6061 stir-cast composite alloys.