Molecular dynamics simulation plays an important role in material simulation, biopharmaceuticals, and other areas. In recent years, the development of AI-for-Science has greatly improved the accuracy of neural network force fields in predicting energy, force, and other properties, compared to traditional methods using potential functions. Neural network force field models may challenges such as hyperparameter settings and gradient explosion when trained by the first-order method. Based on an optimizer named reorganized layer extended Kalman filtering, this study provides several strategies to avoid hyperparameters and offers theoretical evidence for preventing gradient explosion. This study also proposes an alternate training method and analyzes its accuracy gains and time costs. A performance model of block thresholding is proposed, and its effectiveness is explored. Additionally, the property of preventing gradient explosion is proven, and the optimizer’s robustness with respect to activation functions and weight initialization is validated. Four typical neural network force field models are tested on 11 representative datasets. Results show that when the proposed optimizer and the first-order optimizer achieve comparable prediction accuracy, the proposed optimizer is 8 to 10 times faster than the first-order optimizer. It is believed that the proposed training method can inspire other AI-for-Science applications.