Machine Learning Driven Hybrid Surrogate Modeling and NSGA-II Optimization of Machining Performance Parameters

A hybrid surrogate modeling and machine-learning framework was designed for EN31 alloy turning to explore the machining performance in liquid CO 2 assisted minimum quantity lubrication (MQL). The effect of flow rate, cutting velocity and depth of cut on material removal rate (MRR), surface roughness (Ra), rake-face temperature rise (Trake) and cutting force were modeled by using 18 observations obtained experimentally. Models were comparatively evaluated using nested Leave-One-Out Cross-Validation (LOOCV) with model-specific hyperparameter optimization, i.e., Linear Regression, Random Forest, Support Vector Regression (SVR) and Artificial Neural Network (ANN). The updated results show the predictive performance depending on the response. ANN provided the best baseline prediction for MRR and Ra, whereas SVR achieved the best prediction among the baseline models for Trake and cutting forceThe hybrid surrogate model, consisting of second-order polynomial trend and Gaussian Process residual learning, obtained R 2 values of 0.9882, 0.8560, 0.9901 and 0.9877 for MRR, Ra, Trake and cutting force respectively. Sensitivity analysis, predictive uncertainty quantification, and quantitative SHAP analysis were used to increase the interpretability of model predictions. Finally, NSGA-II was used for simultaneous maximization of MRR and minimization of Ra, Trake and cutting force in the experimentally investigated domain of parameters. The Pareto solutions obtained provide alternative machining conditions according to productivity, surface quality, thermal response and mechanical loading requirements. Thus, the proposed framework offers an interpretable, data-efficient approach for predictive modeling and multi-objective decision making in sustainable machining.