📄️ NEP Model
The NEP model was initially implemented in the GPUMD software package (2022b NEP4 GPUMD). NEP training in GPUMD uses the separable natural evolution strategy (SNES), which is simple to implement as it does not rely on gradient information. However, for standard supervised learning tasks, especially deep learning, gradient-based optimization algorithms are more suitable. We implemented the NEP model (NEP4, with the network structure shown in Figure 1) in the PWMLFF 2024.5 version, which can use the gradient-based LKF or ADAM optimizers in PWMLFF for model training.