This paper proposes a neural-network-based method to realize the parameter optimization of permanent magnet synchronous motor model predictive control. Firstly, different combinations of parameters are selected to perform multiple simulations, and the key performance indicators (such as average switching frequency of the inverter, total harmonic distortion, etc.) of the system are extracted. Then, the acquired data are used to train the neural network. The trained neural network performs as a substitute for the simulation model, which can estimate the performance indicators of the system quickly and accurately corresponding to arbitrary combination of parameters. Thirdly, the optimal parameters for three different user-defined fitness functions under different load torque were verified, and the performance indicators predicted by neural network turned out to be close to the results of the simulation model (error less than 5%). It is proved that the neural network can perform as a better substitute for simulation model to realize the fast exhaustive search of the optimal combination of parameters. Finally, the consistency of the chosen parameters was verified by hardware-in-the-loop simulation experiments.