Data-driven prediction and optimization of surface roughness in vibration-assisted ball burnishing of 42CrMo steel

Vibration-assisted ball burnishing (VABB) has demonstrated significant potential in improving the surface quality of high-strength steels by reducing roughness and enhancing functional performance. However, the evolution of surface roughness under multiple interacting parameters remains complex, and reliable prediction with limited experimental data is still challenging. This work establishes a small-sample machine learning framework integrating VABB experiments with predictive optimization to quantitatively reveal the coupling between process parameters and 3D surface morphology of 42CrMo steel. Three approaches, response surface methodology (RSM), Bayesian-optimized support vector machine (BO-SVM), and random forest-enhanced backpropagation neural network (RF-BPNN), were compared in terms of accuracy and generalization. The RF-BPNN model achieved the best performance, with determination coefficients (R 2 ) of 0.9573, 0.7764, and 0.8293 for Sa, Sq, and Spk, respectively, and corresponding RMSE values of 0.0422, 0.1190, and 0.2325. These results indicate that RF-BPNN provides a robust data-driven tool for predicting and optimizing surface roughness in VABB. The proposed framework not only contributes to the understanding of roughness formation mechanisms but also offers practical guidance for process design and performance control of high-strength steel components.