Machine Learning & Interoperability

DeePMD-kit builds deep-learning models of the interatomic potential energy and forces (the “Deep Potential” family of models) from quantum-mechanical reference data, aiming to combine ab initio accuracy with the efficiency of classical MD. It offers a range of descriptors (se_e2_a, se_e3, DPA-2, DPA-3, and others) and can train with TensorFlow, PyTorch, JAX, or Paddle backends. Trained models drive molecular dynamics through LAMMPS, ASE, i-PI, and GROMACS, covering systems from finite molecules to extended metallic and chemically bonded materials. It is developed within the DeepModeling community (source on GitHub).

NequIP trains E(3)-equivariant message-passing neural-network interatomic potentials: by building the rotational, translational, and inversion symmetries of physics directly into the network, it reaches high accuracy with comparatively little training data. Trained models can be deployed for molecular dynamics in LAMMPS (via a dedicated pair style and the ML-IAP interface), the Atomic Simulation Environment (ASE), OpenMM, and torch-sim. It is developed by the MIR group (source on GitHub).

Allegro is a strictly-local E(3)-equivariant deep-learning interatomic potential from the same group as NequIP, and it is built on the same nequip framework. By keeping the model strictly local it scales to very large systems and across many compute nodes while retaining equivariant accuracy. It integrates with LAMMPS through a dedicated pair style/plugin for large parallel molecular dynamics with MPI and GPU acceleration; training and deployment follow the NequIP documentation (source on GitHub).

MACE constructs machine-learning force fields using higher-order equivariant message passing, which captures many-body atomic interactions for fast and accurate predictions. In addition to training your own models, it ships pre-trained “foundation” models that can be used out of the box across the periodic table, and it can provide derived quantities such as analytical Hessians, dipole moments, and polarizabilities. Trained models run in molecular dynamics through ASE, LAMMPS, and OpenMM. It is developed by the ACEsuit community (source on GitHub).

SevenNet (Scalable EquiVariance-Enabled Neural Network) is a graph neural-network interatomic potential whose core model builds on NequIP. Its distinguishing feature is a parallelization scheme designed for LAMMPS that runs GNN potentials efficiently across many GPUs, reaching near-ideal strong scaling and handling systems of well over 100,000 atoms. It ships pretrained universal models (e.g. SevenNet-0, trained on Materials Project data), a fine-tuning workflow, and an ASE calculator. It is developed by MDIL-SNU (source on GitHub).

CHGNet is a pretrained universal (“foundation”) graph neural-network interatomic potential from the Ceder group, trained on the Materials Project Trajectory Dataset (energies, forces, stresses, and magnetic moments from more than 1.5 million structures). It is charge-informed: by predicting site magnetic moments it gives access to charge/oxidation-state information alongside the usual energy and forces. CHGNet runs molecular dynamics natively through its ASE calculator and built-in MD class, so it works across the ASE ecosystem. LAMMPS support is available through AdvanceSoft’s customized LAMMPS variant (an external fork providing a graph-neural-network-potential package, ML-GNNP, with a pair_style chgnet) rather than through mainline LAMMPS (source on GitHub).

The Atomic Simulation Environment (ASE) is a Python library for setting up, manipulating, running, visualizing, and analyzing atomistic simulations. Its unified “calculator” interface connects more than 40 external codes — including LAMMPS, VASP, Quantum ESPRESSO, CP2K, GPAW, and many machine-learning potentials — so a workflow (structure relaxation, molecular dynamics, nudged elastic band, phonons, …) can be scripted once and then run with different engines. This makes ASE a common hub for building complex, multi-code workflows. It is distributed under the GNU LGPL.

OpenKIM (Open Knowledgebase of Interatomic Models) is a curated repository of conventional and machine-learning interatomic potentials, paired with a database that benchmarks each model against a range of material properties (using its Crystal Genome technology) so users can pick a potential suited to their problem. The standardized KIM API lets LAMMPS, ASE, DL_POLY, GULP, and other codes load any archived model in a plug-and-play fashion; it can also wrap a package’s native models alongside the curated reference implementations. OpenKIM is NSF-funded, with a strong emphasis on reproducibility and verified coding integrity.