Prediction of surface roughness in boring of 1.2311 material using machine learning enhanced by virtual sampling methods

Abstract Accurate prediction of surface roughness is essential for quality assurance in boring operations. However, the long and slender structure of boring bars induces dynamic effects that make obtaining consistent surface quality challenging, while extensive experiments are costly and time-consuming. Traditional regression models fall short in predictive performance and cannot adequately capture the nonlinear interactions between cutting parameters. This study presents a machine learning-based methodology to estimate the surface roughness of 1.2311 plastic mold steel during boring on a milling machine. The primary objective is to model the connection between boring variables and hole quality under limited data conditions (72 data points). To overcome insufficient samples, the proposed framework integrates feature augmentation, principal component analysis (PCA), and virtual sampling (VSG) (0 to 500 in increments of 50) with different interpolation techniques. Results indicate that weighted interpolation-based VSG after feature space expansion significantly improves prediction accuracy by strengthening the representation of cutting parameter–Ra relationships. The highest-performing model, combining the Decision Tree algorithm with weighted interpolation and augmented features, achieved a minimum a Root Mean Square Error (RMSE) of 0.28 μm and Mean Absolute Percentage Error (MAPE) of 12.42%. Compared to models trained solely on real experimental data, prediction accuracy improved by 61.9% in MAPE and 34.88% in RMSE, demonstrating the effectiveness of the proposed virtual sampling and feature engineering strategy under limited experimental data conditions. The findings highlight how structured data enhancement can improve surface roughness prediction in milling-based boring and support more reliable process planning.