To achieve efficient and rapid optimization for high torque and low noise permanent magnet synchronous motors, this paper proposes a multi-layer surrogate model-based optimization method for IPMSM (Interior Permanent Magnet Synchronous Motor) based on sensitivity classification of structural parameters. Firstly, using a hybrid model of “FEM + Unit Force Wave Response,” the key order electromagnetic forces causing electromagnetic noise in various operating conditions of the motor are obtained. Their amplitudes, along with the motor’s average output torque and torque ripple, are taken as optimization objectives. By analyzing the sensitivity of structural parameters using the random forest algorithm, the selection and classification of structural parameters are achieved. A hierarchical optimization is then performed using a combination of a multi-island genetic algorithm, a multi-objective particle swarm optimization algorithm, and parameterized scanning. Compared with traditional multi-field coupled optimization methods, this method saves computational resources while reducing calculation time by 54.9%. After optimization, the average output torque is increased by 34.6% compared to before optimization, the amplitude of key order electromagnetic forces of the motor is reduced by 13.7%, and torque ripple is reduced by 67.8%.