Skip to content

Navigation Menu

Sign in
Sign up

v3.2.0b0 #5499

njzjz announced in Announcement
Jun 6, 2026 · 0 comments
Discussion options

Highlights

DPA4/SeZM

We are excited to introduce DPA4/SeZM in DeePMD-kit, bringing the next generation of high-accuracy, high-efficiency machine-learning interatomic potentials to the Deep Modeling community.
DPA4 is designed for the era of Large Atomistic Models (LAMs): it pushes the accuracy-cost frontier by combining strong SE(3)-equivariant modeling power with dramatically reduced training cost. Built around an EMFA SO(2)-equivariant convolution, DPA4 uses edge-conditioned low-rank equivariant products, multi-focus nonlinear message updates, and envelope-gated attention to deliver excellent accuracy without relying on oversized models or massive training budgets.
In benchmark results reported in the DPA4 paper, DPA4-Pro achieves the best Combined Performance Score on Matbench Discovery, while compact DPA4 variants outperform much larger baselines with far fewer parameters and substantially lower training compute. On SPICE-MACE-OFF, DPA4 also sets a new accuracy-cost Pareto frontier for molecular energy and force prediction.
image

Highlights

  • New DPA4/SeZM descriptor for the PyTorch backend of DeePMD-kit.
  • State-of-the-art performance on Matbench Discovery and SPICE-MACE-OFF benchmarks.
  • Much lower training cost compared with previous leading equivariant models.
  • Compiler-friendly training path with native support for torch.compile, enabling up to about 3x wall-clock speedup in reported settings.
  • Efficient SO(2)-equivariant design in local edge frames, reducing the cost of equivariant angular computation.
  • Lebedev-grid projection to preserve SO(3) equivariance through nonlinear operations.
  • Native ZBL short-range repulsion support, improving robustness for close-contact and extreme configurations.

DPA4 shows that top-tier accuracy no longer has to come with top-tier compute cost. With this release, DeePMD-kit makes the new DPA4/SeZM architecture available to researchers and developers who want to train accurate, efficient, and physically consistent interatomic potentials for materials, molecules, and future large atomistic model pretraining.
Learn more in the paper: DPA4: Pushing the Accuracy-Cost Frontier of Interatomic Potentials with EMFA SO(2) Convolution.

New features in models and training strategies

  • DPA3 now supports add_chg_spin_ebd, sequential_update, use_default_pf, and decoupling of charge_spin from fparam, improving its handling of charge and spin information.
  • The property fitting network now supports element-type-wise bias.
  • New loss-related features include force L2 norm loss, MAE loss, and support for three-value Huber delta.
  • The PyTorch backend now includes full validation support, including validation on LMDB-format datasets. Training stability and optimization are also improved through the addition of EMA shadow models, the WSD learning-rate scheduler, and updates to the HybridMuon optimizer. LMDB support was added to the pt/dpmodel data pipeline, including a new LMDB dataloader and max/filter modes.

Beta: The exportable PyTorch backend (pt_expt)

An exportable PyTorch backend (pt_expt) is added based on the Array API, torch.export, and torch.compile. The usage of the Array API makes it accessible to almost all models. The old PyTorch backend will be deprecated in the future, since TorchScript has been deprecated by the PyTorch team. Try the new backend using dp --pt-expt train input.json.

The exportable PyTorch backend received a major expansion in this release, covering new model types, training workflows, evaluation interfaces, data handling, and deployment support. It now supports Linear Energy Model, DeepSpin, multi-task training, missing losses for spin/DOS/tensor/property tasks, and new evaluation APIs such as eval_typeebd, eval_descriptor, and eval_fitting_last_layer. The backend also adds support for dp compress, dp finetune, dp change-bias, .pt training checkpoints in DeepEval, LMDB datasets, pluggable neighbor-list strategies, and efficient O(N) vesin neighbor lists for Python/ASE inference. In addition, .pt2 AOTInductor-based C/C++ inference is now available for DPA1/DPA2/DPA3 models, with improved export/loading tests and multi-rank LAMMPS support for GNN models. Together, these changes make pt_expt much closer to a complete experimental PyTorch backend for training, evaluation, deployment, and production MD workflows.

Beta: Model training in the JAX backend

The JAX backend now supports model training. In the previous versions, it only supported inference. Same as the exportable PyTorch backend, the JAX backend is built on the Array API, making it accessible to almost all models. Try JAX training using dp --jax train input.json.

Agent Skills

Official Agent Skills for DeePMD-kit are now available in the skills directory, introduced as part of this work. Installing these skills empowers your agents to seamlessly interact with DeePMD-kit, enabling them to train Deep Potential models and execute molecular dynamics simulations.

What's Changed

Breaking Changes

New Features

Enhancement

Documentation

Build and release

Bug fixings

Other Changes

Full Changelog: v3.1.3...v3.2.0b0

This discussion was created from the release v3.2.0b0.
You must be logged in to vote

Replies: 0 comments

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
1 participant

AltStyle によって変換されたページ (->オリジナル) /