Explainable Machine Learning Framework for EDM Machinability Analysis of Ti6Al4V and Rex734 Biomaterials

Abstract

This study introduces an explainable ML framework that couples physics-informed feature engineering with SHAP-based interpretability analysis for EDM machinability prediction of two clinically important biomaterials, Ti6Al4V and Rex734, machined under identical conditions. Eighteen experiments (two materials, three current levels, three pulse durations) were conducted, and 12 physics-informed features were engineered from process parameters and material thermal properties. A nested leave-one-out cross-validation protocol with bootstrap confidence intervals was applied to benchmark 13 classifiers and 15 regressors. For material classification, Linear Discriminant Analysis, Logistic Regression, and SVM-Linear each achieved 94.44% accuracy (95% CI: 83.33-100.00%, MCC = 0.894), confirmed significant by permutation testing (p = 0.005). For regression, SVR-RBF predicted surface roughness with R-2 = 0.928 and MAPE = 5.87%, while CatBoost predicted recast layer thickness with R-2 = 0.884. SHAP analysis identified surface roughness (33.2%) and pulse duration (23.8%) as the dominant classification features, and revealed that material thermal conductivity and pulse duration jointly govern roughness formation-findings that align quantitatively with two-way ANOVA, where pulse duration explained 90% of Ra variance for Ti6Al4V. A physics-aware data augmentation strategy expanded the dataset from 18 to 90 samples and improved mean classification accuracy from 0.829 to 0.991 (Wilcoxon p = 0.0002), although these augmented accuracies represent upper-bound estimates due to indirect data leakage in the evaluation protocol, and the direction rather than the absolute magnitude of improvement constitutes the robust finding. The mutual validation between ANOVA and SHAP provides a methodological template for small-sample manufacturing studies in which statistical and ML approaches reinforce each other.