DeePMD-kit

Overview

DeePMD-kit connects atomistic reference data, machine-learning potential development, and molecular-dynamics deployment. It is a toolkit rather than a single predictive model: users can adapt pretrained DPA4 checkpoints or configure models for training from scratch. The documented scope includes finite molecules, periodic solids, covalent systems, and metals, with workflows for single-task, multi-task, and distributed training.

Training and fine-tuning use target reference data in DeePMD's NumPy format; dpdata provides a route for converting structures and trajectories. Depending on the selected model, predictions can include energies, atomic forces, virials, Hessians, magnetic forces, dipoles, polarizabilities, and other physical properties. The workflow proceeds through testing and model export to Python or native inference interfaces and supported simulation engines, including LAMMPS and i-PI. DPA-ADAPT offers a separate path for reusing supported pretrained representations in downstream property prediction.

Model selection affects deployment constraints. The README presents DPA4 as an accuracy-oriented architecture and DPA4C as a throughput-oriented alternative, without making either universally preferable. DPA4 uses PyTorch, while DPA4C uses the PyTorch Exportable backend; its compressed CUDA inference path requires float32. More broadly, backend and interface support vary by model and feature. The supplied pretrained OMat24 checkpoints target inorganic materials in that chemical space, so accuracy outside it requires validation. These constraints make target-domain evaluation and compatibility checks important before simulation deployment.

Key Features

• Full-model fine-tuning of pretrained DPA4 checkpoints, alongside training from scratch with single-task, multi-task, and distributed workflows.
• Model-dependent prediction of energies, forces, virials, Hessians, spin-related quantities, dipoles, polarizabilities, density of states, and custom properties.
• Support for selected models across TensorFlow, PyTorch, JAX, and Paddle, with conversion paths for compatible architectures.
• Testing, freezing, embedding extraction, and supported model-compression and export paths, including AOTInductor .pt2 export.
• Python, C, C++, and Node.js inference interfaces, plus documented simulation integrations such as LAMMPS, i-PI, and ASE.
• Hybrid potentials and physics corrections, including DPLR electrostatics, DPRc range correction for QM/MM, and analytical ZBL bridging.

Use Cases

• Intended evaluation: fine-tune an OMat24 DPA4 checkpoint on target inorganic-material reference data and assess its accuracy before molecular-dynamics deployment.
• Intended evaluation: use the supplied water examples to compare DPA4 and DPA4C for a workload's accuracy, throughput, and deployment requirements.
• Intended evaluation: reuse supported pretrained representations through DPA-ADAPT for downstream physical-property prediction.
• Intended evaluation: assess a compatible spin-aware or response-property model for magnetic-force, dipole, or polarizability prediction.

How to Use

1. Consult the installation guide and choose a documented installation route for your backend and hardware. The linked sources require Python 3.10 or later.
2. Use the model guide to select a physical target and compatible architecture. Decide between a pretrained starting point and training from scratch; check backend and deployment restrictions.
3. Prepare target training and validation data in DeePMD's NumPy format, using the dpdata guidance when converting structures or trajectories.
4. Follow the fine-tuning guide or quick-start notebook. For pretrained DPA4 adaptation, use the matching released configuration and preserve its complete model section and type_map.
5. Evaluate target-domain predictions using the testing guidance, then consult the export guidance and integration hub before deployment.