Deep Learning Framework With Optuna-Based Hyperparameter Tuning for Predicting Dry Turning Process Performance of 42CrMo4 Steel

Abstract

This study presents an integrated experimental and data-driven modeling framework for predicting key machining responses such as tool wear (TW), material removal rate (MRR), and surface roughness (Ra) during dry turning of 42CrMo4 alloy steel. Cylindrical workpieces (φ 40 mm × 50 mm) were machined using CVD-coated TNMG160308-MT-TT8115 inserts on a CNC lathe under controlled environmental conditions. The experimental plan followed a L27 orthogonal array by varying three input parameters: cutting speed (Vc: 1000-1400 rpm), feed rate (fz: 0.15-0.25 mm/rev), and depth of cut (ap: 0.5-1.0 mm). To model the nonlinear relationship between inputs and responses, Deep Feedforward Neural Networks (DFNNs) were optimized using three techniques: grid search, genetic algorithm (GA) and Optuna-based hyperparameter optimization. Among them, the Optuna-DFNN demonstrated superior performance with significantly lower mean squared error (MSE) values across all outputs for TW: 1.1071× 10-5, MRR: 836,330 and Ra: 1.1037 as compared to grid search (TW: 3.2247× 10-5, MRR: 1,981,800, Ra: 1.4782) and GA (TW: 4.4856× 10-5, MRR: 4,326,800, Ra: 1.4806). Regression analysis of the Optuna-DFNN tool wear model yielded an R2 of 0.818 and Pearson's r of 0.904, indicating strong predictive correlation. The moderately high R2 value is attributed to limited size and inherent variability of the experimental dataset, which posed challenges in capturing full complexity of the machining process. The relative prediction error remained below 5% for most test cases, validating the model's generalization capability. These findings confirm that Optuna-based DFNN, with intelligent hyperparameter tuning, enables more accurate learning of process dynamics than conventional or heuristic optimization methods, making it a powerful tool for predictive modeling and real-time control in smart manufacturing systems.