Abstract
Ultrasonic rolling is an effective technique for enhancing surface integrity, and surface integrity is closely related to fatigue performance. The process parameters of ultrasonic rolling critically affect the improvement of surface integrity. This study proposes an optimization method for process parameters by combining machine learning (ML) with the NSGA-II. Five ML models were trained to establish relationships between process parameters and surface residual stress, hardness, and surface roughness by incorporating feature augmentation and physical information. The best-performing model was selected and integrated with NSGA-II for multi-objective optimization. Ultrasonic rolling tests based on a uniform design were performed, and a dataset was established. The objective was to maximize surface residual stress and hardness while minimizing surface roughness. For test specimens with an initial surface roughness of 0.54 µm, the optimized process parameters were a static pressure of 900 N, a spindle speed of 75 rpm, a feed rate of 0.19 mm/r, and rolling once. Using optimized parameters, the surface residual stress reached -920.60 MPa, surface hardness achieved 958.23 HV, surface roughness reduced to 0.32 µm, and contact fatigue life extended to 3.02 × 10(7) cycles, representing a 52.5% improvement compared to untreated specimens and an even more significant improvement over without parameter optimization.