AI-Based Modeling and Analysis of Surface Roughness, Tool Wear, and Material Removal Rate in Dry Hard Turning of Skd11 for Adaptive Control Insights

Dry hard turning of heat-treated SKD11 tool steel offers a sustainable alternative to grinding but poses significant challenges in maintaining surface integrity and tool life due to severe wear and high cutting temperatures. This study develops a physically interpretable, AI-assisted modeling framework that integrates empirical and machine learning methods to predict and optimize surface roughness ([Formula: see text]), tool wear ([Formula: see text]), and material removal rate (MRR). Seventeen controlled experiments were conducted under varying cutting speed ([Formula: see text]), feed rate ([Formula: see text]), and depth of cut ([Formula: see text]) conditions using CBN inserts (TNGA160404–CBN1020) in dry cutting. A hierarchical modeling approach was established, where [Formula: see text] acts as an intermediate variable linking process parameters to [Formula: see text], while MRR was computed analytically. The Response Surface Methodology (RSM) provided the baseline regression models with [Formula: see text] for [Formula: see text] and [Formula: see text] for [Formula: see text] when wear was included. Support Vector Regression (SVR) further enhanced prediction accuracy to [Formula: see text], demonstrating strong nonlinear capability and physical consistency. Multi-objective optimization using the MOPSO–TOPSIS framework achieved a balanced solution at [Formula: see text]–160[Formula: see text]m/min, [Formula: see text]–0.07[Formula: see text]mm/rev, and [Formula: see text]–0.15[Formula: see text]mm, resulting in approximately 10% reduction in surface roughness, 12% reduction in tool wear, and 8% improvement in productivity compared with baseline conditions. The proposed framework bridges physical modeling and AI-based prediction, offering a robust, explainable tool for adaptive control in precision hard turning and a reliable foundation for intelligent machining under limited-data environments.