DeePMD-kit trains deep learning models to represent potential energy surfaces and atomic forces for molecular dynamics simulations. It interfaces with multiple ML backends (TensorFlow, PyTorch, JAX, Paddle) for automatic training and with classical MD codes like LAMMPS and i-PI for efficient simulation. The project addresses the accuracy-versus-efficiency trade-off in computational chemistry and materials science, applicable to molecular systems, extended systems, and metallic or chemically bonded materials.