Machine Learning-Driven Optimization of Energy Efficiency in Marble Computer Numerical Control Machining for Sustainable Manufacturing

Abstract This study investigates the specific energy (Se) and material removal rate (MRR) during the computer numerical control (CNC) machining of marble using three toolpath strategies: external lines, linear, and spiral across varying cutting depths and feed rates. Analysis of variance (ANOVA) confirmed statistically significant differences in Se across toolpath types, cutting depths, and MRR levels (p < 0.05). Among the strategies, external lines exhibited the highest energy consumption, and deeper cuts (2.0 mm) were associated with increased Se values. Furthermore, higher MRR levels were strongly correlated with reduced Se (p < 0.001), highlighting their role in energy-efficient machining. Correlation analysis revealed strong linear relationships between Se and both cutting depth (R2 = 0.70) and feed rate (R2 = 0.70), while the correlation with MRR was relatively weak (R2 = 0.16), suggesting a more complex or indirect relationship. Feature importance analysis using the XGBoost algorithm identified MRR as the most influential predictor, contributing 96.05% to the model's predictive accuracy. A series of machine learning models were developed to forecast Se under varying machining conditions, with CatBoost, LightGBM, and XGBoost demonstrating the highest predictive performance (R2 > 0.98) and generalization ability. Among the evaluated models, CatBoost yielded the best performance (R2 = 0.983), followed closely by LightGBM (R2 = 0.983) and XGBoost (R2 = 0.982), demonstrating high predictive accuracy and minimal overfitting. In terms of toolpath strategies, external lines produced the most accurate Se predictions, followed by linear and spiral trajectories. In addition, a K-means-based clustering approach was used to classify specific energy levels based on cutting force and energy metrics. Gradient boosting achieved the highest classification accuracy and precision (both 0.75), as well as the highest AUC scores across all ROC analyses, confirming its robustness in energy-based categorization. The study confirms that MRR, cutting depth, and toolpath strategy are the main determinants of energy consumption, with ensemble-based ML models offering both high accuracy and strong generalization. SHapley Additive exPlanations (SHAP)-supported interpretability enables transparent parameter optimization for adaptive and efficient CNC operations. These insights contribute to the development of intelligent machining systems by integrating interpretable, data-driven approaches into sustainable marble processing workflows.