Explainable Machine Learning with Optimized Tree-Based Models and Statistical Validation for Data-Driven Wear...

Accurate prediction of wear behavior in steel materials is essential for enhancing component durability and optimizing manufacturing processes. This study presents a comparative data-driven framework for wear prediction using Decision Tree (DT), Random Forest (RF), and a Physics-Informed Neural Network (PINN). Hyperparameter optimization for the DT and RF models was performed using GridSearchCV with 5-fold cross-validation to obtain optimal model configurations. Model performance was evaluated using the Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and coefficient of determination (R²). In addition, sensitivity analysis was conducted by varying train-test split ratios, while residual analysis, confidence interval estimation, and SHAP-based feature importance analysis were employed to assess model reliability and interpretability. The Random Forest model achieved the best predictive performance on the test dataset with RMSE = 0.00495, MAE = 0.00273, and R² = 0.7244, followed by the Decision Tree model with RMSE = 0.00565, MAE = 0.00275, and R² = 0.6406. The implemented Physics-Informed Neural Network (PINN), incorporating a load-time physics constraint, exhibited significantly poorer performance, yielding RMSE = 0.15872, MAE = 0.12599, and R² = -282.62, indicating that the available experimental dataset and simplified physics constraint were insufficient for effective physics-informed learning. SHAP analysis identified surface condition, material type, sliding time, and the load-time interaction as the most influential parameters governing wear behavior. The findings demonstrate that optimized tree-based machine learning models, particularly the Random Forest model, provide accurate, interpretable, and computationally efficient predictions for wear estimation in the investigated steel materials.