Models

MLIP Arena provides a unified interface to 15+ foundation machine learning interatomic potentials. Every model is exposed as an ASE Calculator and enumerated in MLIPEnum for easy iteration across benchmarks.

MLIPEnum

MLIPEnum is a Python Enum built dynamically at import time from registry.yaml. Each member’s name is the model’s display name and its value is the calculator class.

from mlip_arena.models import MLIPEnum

# Iterate over all available models
for model in MLIPEnum:
    print(model.name, model.value)  # e.g. "MACE-MP(M)" <class MACE_MP_Medium>

# Access a specific model by name
model_class = MLIPEnum["MACE-MP(M)"].value
calc = model_class()  # instantiates the ASE Calculator

MLIPEnum only contains models whose packages are installed in the current environment. If a model’s dependency is missing, it is silently skipped with a warning rather than raising an import error.

The MLIP base class

HuggingFace-hosted models inherit from MLIP, defined in mlip_arena/models/__init__.py:

# mlip_arena/models/__init__.py (lines 60–103)
class MLIP(
    nn.Module,
    PyTorchModelHubMixin,
    tags=["atomistic-simulation", "MLIP"],
):
    def __init__(self, model: nn.Module) -> None:
        super().__init__()
        self.model = model

    @classmethod
    def from_pretrained(
        cls,
        pretrained_model_name_or_path: str | Path,
        **model_kwargs,
    ) -> Self:
        return super().from_pretrained(
            pretrained_model_name_or_path,
            **model_kwargs,
        )

    def forward(self, x):
        return self.model(x)

The PyTorchModelHubMixin inheritance means any MLIP subclass can be uploaded to and downloaded from HuggingFace Hub with push_to_hub() and from_pretrained().

MLIPCalculator

MLIPCalculator combines MLIP (the neural network) with ASE’s Calculator interface:

# mlip_arena/models/__init__.py (lines 105–168)
class MLIPCalculator(MLIP, Calculator):
    implemented_properties: list[str] = ["energy", "forces", "stress"]

    def __init__(
        self,
        model: nn.Module,
        device: torch.device | None = None,
        cutoff: float = 6.0,
        calculator_kwargs: dict = {},
    ):
        MLIP.__init__(self, model=model)
        Calculator.__init__(self, **calculator_kwargs)
        self.device = device or get_freer_device()
        self.cutoff = cutoff
        self.model.to(self.device)

    def calculate(
        self,
        atoms: Atoms,
        properties: list[str],
        system_changes: list = all_changes,
    ):
        data = collate_fn([atoms], cutoff=self.cutoff).to(self.device)
        output = self.forward(data)
        self.results = {}
        if "energy" in properties:
            self.results["energy"] = output["energy"].squeeze().item()
        if "forces" in properties:
            self.results["forces"] = output["forces"].squeeze().cpu().detach().numpy()
        if "stress" in properties:
            self.results["stress"] = output["stress"].squeeze().cpu().detach().numpy()

Device selection is automatic: get_freer_device() picks the CUDA GPU with the most free memory, falls back to MPS on Apple Silicon, and finally to CPU.

External ASE calculators vs HuggingFace models

External ASE calculators
HuggingFace models

Most models in the registry are implemented as external ASE calculators under mlip_arena/models/externals/. They wrap third-party packages and set module: externals in the registry.When to use this approach:

The model already ships its own inference code (e.g., mace-torch, chgnet, matgl).
You want to add a model quickly without implementing a custom graph network.

# registry.yaml entry for an external model
MACE-MP(M):
  module: externals
  class: MACE_MP_Medium
  family: mace-mp
  package: mace-torch==0.3.9

Remove any unnecessary entries from the results class attribute on your calculator. Extra properties (e.g., magnetic moments) can cause errors during MD simulations. See the CHGNet implementation as a reference.

Models hosted on HuggingFace inherit from MLIP (which includes PyTorchModelHubMixin) and upload their checkpoints to a HF model repository.When to use this approach:

You want to contribute a new MLIP to the arena with a public checkpoint.
You need version control and reproducibility for your model weights.

Inherit PyTorchModelHubMixin

Add PyTorchModelHubMixin (or the MLIP base class) to your model class definition.

Push to HuggingFace Hub

Create a HuggingFace model repository and upload weights:

model.push_to_hub("your-username/your-model-name")

Implement the I/O interface

Follow the template in mlip_arena/models/README.md to implement the graph construction and calculate() method.

Update registry.yaml

Add your model’s metadata to mlip_arena/models/registry.yaml.

The model registry.yaml structure

Each entry in mlip_arena/models/registry.yaml has the following fields:

# Example entry from registry.yaml
CHGNet:
  module: externals          # Python subpackage under mlip_arena/models/
  class: CHGNet              # Class name to import
  family: chgnet             # Module file name (family.py)
  package: chgnet==0.3.8     # Required pip package
  checkpoint: v0.3.0         # Checkpoint identifier
  username: cyrusyc          # HuggingFace username
  datasets:
    - MPTrj                  # Training datasets
  gpu-tasks:                 # Benchmark tasks this model runs on GPU
    - homonuclear-diatomics
    - stability
    - combustion
    - eos_bulk
    - wbm_ev
  prediction: EFSM           # E=Energy, F=Forces, S=Stress, M=Magmoms
  nvt: true                  # Supports NVT ensemble
  npt: true                  # Supports NPT ensemble
  doi: https://doi.org/10.1038/s42256-023-00716-3
  date: 2023-02-28
  license: BSD-3-Clause

Field	Description
`module`	Subpackage under `mlip_arena/models/` (`externals` or `huggingface`)
`class`	Python class name to import and register
`family`	File name of the module containing the class
`package`	Exact pip package and version required
`checkpoint`	Model checkpoint identifier (path, name, or version string)
`datasets`	Training datasets used
`gpu-tasks`	Benchmark task IDs this model participates in on GPU
`cpu-tasks`	Benchmark task IDs this model participates in on CPU
`prediction`	Prediction types: `E` energy, `F` forces, `S` stress, `M` magnetic moments
`nvt` / `npt`	Whether NVT and NPT molecular dynamics are supported
`doi`	Publication DOI
`license`	Model license

Dynamic loading at import time

When you import mlip_arena.models, the __init__.py reads registry.yaml and dynamically imports each model class:

# mlip_arena/models/__init__.py (lines 35–56)
with open(Path(__file__).parent / "registry.yaml", encoding="utf-8") as f:
    REGISTRY = yaml.safe_load(f)

MLIPMap = {}
for model, metadata in REGISTRY.items():
    try:
        module = importlib.import_module(
            f"{__package__}.{metadata['module']}.{metadata['family']}"
        )
        MLIPMap[model] = getattr(module, metadata["class"])
    except (ModuleNotFoundError, AttributeError, ImportError, Exception) as e:
        logger.warning(e)
        continue

MLIPEnum = Enum("MLIPEnum", MLIPMap)
logger.info(f"Successfully loaded models: {list(MLIPEnum.__members__.keys())}")

Models whose packages are not installed produce a warning and are excluded from MLIPEnum. This allows you to install only a subset of models and still use the library.

Available models

Model	Family	Training datasets	Predictions	NVT	NPT
MACE-MP(M)	mace-mp	MPTrj	EFS	✓	✓
MACE-MPA	mace-mp	MPTrj, Alexandria	EFS	✓	✓
CHGNet	chgnet	MPTrj	EFSM	✓	✓
M3GNet	matgl	MPF	EFS	✓	✓
MatterSim	mattersim	MPTrj, Alexandria, Proprietary	EFS	✓	✓
ORBv2	orb	MPTrj, Alexandria	EFS	✓	✓
SevenNet	sevennet	MPTrj	EFS	✓	✓
eqV2(OMat)	fairchem	OMat, MPTrj, Alexandria	EFS	✓	✗
eSEN	fairchem	OMat, MPTrj, Alexandria	EFS	✓	✓
EquiformerV2(OC22)	equiformer	OC22	EF	✓	✗
EquiformerV2(OC20)	equiformer	OC20	EF	✓	✗
eSCN(OC20)	escn	OC20	EF	✓	✗
MACE-OFF(M)	mace-off	SPICE	EFS	✓	✓
ANI2x	ani	—	EFS	✓	✓
ALIGNN	alignn	MP22	EFS	✓	✓
DeepMD	deepmd	MPTrj	EFS	✓	✓
ORB	orb	MPTrj, Alexandria	EFS	✓	✓

Prediction types: E = energy, F = forces, S = stress, M = magnetic moments.

Using get_calculator()

get_calculator() from mlip_arena.tasks.utils is the recommended way to instantiate a calculator. It handles device selection, optional dispersion correction via TorchDFTD3Calculator, and accepts multiple input types:

# mlip_arena/tasks/utils.py (lines 56–122)
from mlip_arena.tasks.utils import get_calculator
from mlip_arena.models import MLIPEnum
from ase import units

# From MLIPEnum member
calc = get_calculator(MLIPEnum["MACE-MP(M)"])

# From string name
calc = get_calculator("CHGNet")

# With dispersion correction (DFT-D3)
calc = get_calculator(
    MLIPEnum["MACE-MP(M)"],
    dispersion=True,
    dispersion_kwargs=dict(
        damping="bj", xc="pbe", cutoff=40.0 * units.Bohr
    ),
)

# Pass a custom ASE Calculator class directly
from ase.calculators.emt import EMT
calc = get_calculator(EMT, calculator_kwargs={"asap_cutoff": True})

get_calculator() returns a BaseCalculator instance, so it is compatible with all MLIP Arena tasks.

Get Started

Core Concepts

Tasks

Benchmarks

Contributing

MLIPEnum

The MLIP base class

MLIPCalculator

External ASE calculators vs HuggingFace models

The model registry.yaml structure

Dynamic loading at import time

Available models

Using get_calculator()

Get Started

Core Concepts

Tasks

Benchmarks

Contributing

​MLIPEnum

​The MLIP base class

​MLIPCalculator

​External ASE calculators vs HuggingFace models

​The model registry.yaml structure

​Dynamic loading at import time

​Available models

​Using get_calculator()

MLIPEnum

The MLIP base class

MLIPCalculator

External ASE calculators vs HuggingFace models

The model registry.yaml structure

Dynamic loading at import time

Available models

Using get_calculator()