Interpretable machine learning models for predicting corrosion behaviour in Titanium alloys

Abstract

Titanium alloys are prized for their exceptional corrosion resistance, yet controlling their corrosion behaviour remains challenging due to the complex interactions between alloying elements and corrosive media. This study developed a machine learning model to predict the corrosion current density and potential using data from 115 titanium alloys. The results indicate that the random forest (RF) model, using alloy composition and nine environmental factors as variables, is the most effective for predicting corrosion. The RF model outperforms others in predicting corrosion current density and potential on the test set, achieving the highest R² values (0.670 and 0.489) and the lowest mean absolute error values (0.367 and 0.107). Sixteen physical and chemical features and nine environmental features were used to create new features. Feature reduction results show that the first six principal components were extracted, with their cumulative variance contribution reaching 90.466%, surpassing the 90% threshold. SHAP analysis reveals that Icorr is a more sensitive and direct indicator of corrosion performance, with pH having the greatest impact (coefficient of 0.569). Overall, this study provides valuable insights into optimising titanium alloy compositions for specific corrosive environments through advanced predictive modelling and data-driven analysis.