To address the problems of insufficient geometric freedom and difficulty in balancing multi-physics requirements in the lightweight design of rotor laminations for traditional interior permanent magnet synchronous motors, an optimization design method integrating parametric geometry, sampling under complex constraints, and machine learning was proposed. First, a lightening cavity contour was constructed using two connected second-order Bézier curves defined by five control points, achieving a geometric description with few parameters and high flexibility. Second, an experimental design method integrating acceptance-rejection sampling and a parameter bias strategy was proposed to generate 100 sets of high-quality samples under complex constraints. Then, a surrogate model based on feature engineering, kernel density estimation, and an adaptive weighting strategy was constructed to perform high-precision prediction of average torque, torque ripple, and maximum equivalent stress, with a coefficient of determination R² > 0.89. Finally, a multi-objective optimization scheme was established and solved using a differential evolution algorithm. Finite element simulations indicate that, compared with the lightweight design with traditional circular cavities, the optimized scheme increases the lightening cavity area by 174.6% and reduces the maximum equivalent stress by 18.2% while maintaining electromagnetic performance almost unchanged. This achieves synergistic optimization of multi-physics fields and lightweighting, helps to improve motor economy and power density, and provides a new approach for low-inertia and fast-response motor design.