TorchMD-NET provides neural network architectures for predicting molecular potential energies and forces, which are critical for molecular dynamics simulations. It implements several equivariant neural network architectures including TensorNet, Equivariant Transformers, and Graph Neural Networks, all exposed as PyTorch modules. The trained potentials integrate with GPU-accelerated molecular dynamics codes like ACEMD, OpenMM, and TorchMD for running simulations.