In safety-critical control for permanent magnet synchronous motors (PMSMs), overshooting after adding a spontaneous load is a crucial metric, leading to the unexpected motion of driving equipment, which induces potential unsafe problems. Therefore, it is necessary to develop a control method that effectively reduces overshoot in PMSMs. Recognizing the nature of overshoot effects, a data-driven approach, Gaussian process regression (GPR), is employed to generate the prediction. With a focus on maintaining the advantage of the GPR method, while preserving the physical properties of PMSM, an overshoot reduction-inspired motor physics embedded Gaussian Process Regression method (OR-MPE-GPR) is proposed. Inspired by the shape of the overshoot, the squared exponential (SQE) kernel function is chosen for GPR. Furthermore, by using sufficient conditions to achieve stability, the dynamic stable range and static stable range of updating rate are derived to guarantee the stability of the proposed machine learning control algorithm. Finally, comprehensive simulations and experiments compared with the state-of-the-art methods are conducted, showcasing the superior performance of the proposed method in reducing overshoot while preserving static performance within a stable region.