DeePMD-kit

DeePMD-kit is open-source software for molecular and materials researchers who want to train and run AI interatomic potentials on their own hardware. It learns atomic interactions from quantum-mechanical reference data, then uses those models in molecular dynamics simulations of molecules, solids and metals. It runs on CPUs and GPUs, with support for CUDA and ROCm source builds, and carries the LGPL-3.0-or-later license.

Researchers can adapt pretrained DPA models to their system or train a potential from scratch. Full-model fine-tuning and DPA-4 LoRA are available, with LoRA limited to PyTorch single-task training. DeepPot-SE provides an established baseline, while attention-based DPA models and the SO(3)-equivariant DPA-4 offer other approaches to representing atomic environments. DPA-ADAPT reuses learned representations for downstream property prediction.

The physical targets extend beyond energies and forces to spin, dipoles, polarizabilities and electronic density of states. Models can combine learned interactions with long-range electrostatics or analytical short-range corrections. TensorFlow, PyTorch, JAX and Paddle backends support compatible models, with conversion paths between supported architectures.

Workloads range from laptop fine-tuning to distributed training and MPI-parallel simulations. Compression can reduce inference time and memory use for supported models. Integrations include LAMMPS, i-PI, ASE and GROMACS, alongside Python, C and C++ interfaces for embedding models in scientific applications.