Machine learning-based prediction of surface roughness and machinability in heat-treated Ni–Cr–Mo low-alloy steels

Nickel–chromium–molybdenum (Ni–Cr–Mo) low-alloy steels are commonly used in automotive and aerospace industries for their strength and fracture toughness. However, these steels are difficult to machine. Their poor machinability often results in rapid tool wear, a poor surface finish, and higher production costs. To address these challenges, this study integrates tailored and novel heat treatment processes with machine learning models to predict machining performance. Two heat treatment procedures were examined: conventional isothermal annealing (ITA), which generates a heterogeneous pearlite–ferrite–bainite structure, and a novel multi-step process [Normalized Subcritical Annealing (NSA)], designed to refine the microstructure into a uniform pearlite–ferrite–carbide matrix with reduced hardness variation. Mechanical properties, including yield strength, ultimate tensile strength, elongation, reduction in area, proof stress, Young's modulus, and impact energy, were measured and used as input features. Several regression models, including random forest, XGBoost, ridge, support vector regression, K-nearest neighbors, a stacking regressor, and a multilayer perceptron, were trained to predict two machining outcomes: surface roughness (Ra) and machinability index (%). The results indicate that impact energy and yield strength serve as the most influential features, establishing a direct correlation between toughness and strength in relation to machining performance. Heat treatment-dependent trends were observed. For ITA samples, random forest captured the machinability index more effectively, reflecting the broader property distributions of the heterogeneous microstructure. For NSA samples, random forest provided better predictions of surface roughness “Ra,” consistent with the uniformity of the refined structure. However, the coefficient of determination (R2) was moderate. The models demonstrated high predictive precision with strong industrial relevance, as evidenced by their low error metrics. This accuracy can often be more valuable for industrial decision-making than simply achieving a perfect fit. This study is the first to establish a data-driven link between a novel NSA heat treatment and machine learning predictions for machining outcomes. Overall, NSA treatment is best suited when a consistent surface finish is required, while ITA provides flexibility in tuning machinability, providing practical guidelines for industrial applications. The findings provide a framework that supports alloy design and optimizes machining processes, reducing reliance on trial-and-error methods.