⚔️ MLIP Arena ⚔️

Foundation machine learning interatomic potentials (MLIPs), trained on extensive databases containing millions of density functional theory (DFT) calculations, have revolutionized molecular and materials modeling, but existing benchmarks suffer from data leakage, limited transferability, and an over-reliance on error-based metrics tied to specific density functional theory (DFT) references.

We introduce MLIP Arena, a unified benchmark platform for evaluating foundation MLIP performance beyond conventional error metrics. It focuses on revealing the physical soundness learned by MLIPs and assessing their utilitarian performance agnostic to underlying model architecture and training dataset.

By moving beyond static DFT references and revealing the important failure modes of current foundation MLIPs in real-world settings, MLIP Arena provides a reproducible framework to guide the next-generation MLIP development toward improved predictive accuracy and runtime efficiency while maintaining physical consistency.

MLIP Arena leverages modern pythonic workflow orchestrator 💙 Prefect 💙 to enable advanced task/flow chaining and caching.

[!NOTE] Contributions of new tasks through PRs are very welcome! If you're interested in joining the effort, please reach out to Yuan at cyrusyc@berkeley.edu. See project page for some outstanding tasks, or propose new feature requests in Discussion.

Announcement

• [April 8, 2025] 🎉 MLIP Arena is accepted as an ICLR AI4Mat Spotlight! 🎉 Huge thanks to all co-authors for their contributions!

Installation

From PyPI (prefect workflow only, without pretrained models)

bash pip install mlip-arena

From source (with integrated pretrained models, advanced)

[!CAUTION] We strongly recommend clean build in a new virtual environment due to the compatibility issues between multiple popular MLIPs. We provide a single installation script using uv for minimal package conflicts and fast installation!

[!CAUTION] To automatically download farichem OMat24 checkpoint, please make sure you have gained downloading access to their HuggingFace model repo (not dataset repo), and login locally on your machine through huggginface-cli login (see HF hub authentication)

Linux

```bash

(Optional) Install uv, way faster than pip, why not? :)

curl -LsSf https://astral.sh/uv/install.sh | sh source $HOME/.local/bin/env

git clone https://github.com/atomind-ai/mlip-arena.git cd mlip-arena

One script uv pip installation

bash scripts/install.sh ```

[!TIP] Sometimes installing all compiled models takes all the available local storage. Optional pip flag --no-cache could be uesed. uv cache clean will be helpful too.

Mac

```bash

(Optional) Install uv

curl -LsSf https://astral.sh/uv/install.sh | sh source $HOME/.local/bin/env

One script uv pip installation

bash scripts/install-macosx.sh ```

Quickstart

The first example: Molecular Dynamics

Arena provides a unified interface to run all the compiled MLIPs. This can be achieved simply by looping through MLIPEnum:

```python from mliparena.models import MLIPEnum from mliparena.tasks import MD from mliparena.tasks.utils import getcalculator

from ase import units from ase.build import bulk

atoms = bulk("Cu", "fcc", a=3.6) * (5, 5, 5)

results = []

for model in MLIPEnum: result = MD( atoms=atoms, calculator=getcalculator( model, calculatorkwargs=dict(), # passing into calculator dispersion=True, dispersionkwargs=dict( damping='bj', xc='pbe', cutoff=40.0 * units.Bohr ), # passing into TorchDFTD3Calculator ), # compatible with custom ASE Calculator ensemble="nve", # nvt, nvt available dynamics="velocityverlet", # compatible with any ASE Dynamics objects and their class names totaltime=1e3, # 1 ps = 1e3 fs time_step=2, # fs ) results.append(result) ```

🚀 Parallelize Benchmarks at Scale

To run multiple benchmarks in parallel, add .submit before the task function and wrap all the tasks into a flow to dispatch the tasks to worker for concurrent execution. See Prefect Doc on tasks and flow for more details.

```python ... from prefect import flow

@flow def runalltasks:

futures = []
for model in MLIPEnum:
    future = MD.submit(
        atoms=atoms,
        ...
    )
    future.append(future)

return [f.result(raise_on_failure=False) for f in futures]

```

For a more practical example using HPC resources, please now refer to MD stability benchmark.

List of implemented tasks

The implemented tasks are available under mlip_arena.tasks.<module>.run or from mlip_arena.tasks import * for convenient imports (currently doesn't work if phonopy is not installed).

• OPT: Structure optimization
• EOS: Equation of state (energy-volume scan)
• MD: Molecular dynamics with flexible dynamics (NVE, NVT, NPT) and temperature/pressure scheduling (annealing, shearing, etc)
• PHONON: Phonon calculation driven by phonopy
• NEB: Nudged elastic band
• NEBFROMENDPOINTS: Nudge elastic band with convenient image interpolation (linear or IDPP)
• ELASTICITY: Elastic tensor calculation

Contribute and Development

PRs are welcome. Please clone the repo and submit PRs with changes.

To make change to huggingface space, fetch large files from git lfs first and run streamlit:

git lfs fetch --all git lfs pull streamlit run serve/app.py

Add new benchmark tasks (WIP)

[!NOTE] Please reuse, extend, or chain the general tasks defined above

Add new MLIP models

If you have pretrained MLIP models that you would like to contribute to the MLIP Arena and show benchmark in real-time, there are two ways:

External ASE Calculator (easy)

1. Implement new ASE Calculator class in mlip_arena/models/externals.
2. Name your class with awesome model name and add the same name to registry with metadata.

[!CAUTION] Remove unneccessary outputs under results class attributes to avoid error for MD simulations. Please refer to CHGNet as an example.

Hugging Face Model (recommended, difficult)

1. Inherit Hugging Face ModelHubMixin class to your awesome model class definition. We recommend PytorchModelHubMixin.
2. Create a new Hugging Face Model repository and upload the model file using pushtohub function.
3. Follow the template to code the I/O interface for your model here.
4. Update model registry with metadata

Citation

If you find the work useful, please consider citing the following:

bibtex @inproceedings{ chiang2025mlip, title={{MLIP} Arena: Advancing Fairness and Transparency in Machine Learning Interatomic Potentials through an Open and Accessible Benchmark Platform}, author={Yuan Chiang and Tobias Kreiman and Elizabeth Weaver and Ishan Amin and Matthew Kuner and Christine Zhang and Aaron Kaplan and Daryl Chrzan and Samuel M Blau and Aditi S. Krishnapriyan and Mark Asta}, booktitle={AI for Accelerated Materials Design - ICLR 2025}, year={2025}, url={https://openreview.net/forum?id=ysKfIavYQE} }