⚔️ 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 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.

• Announcement
• Installation
  • From PyPI (prefect workflow only, without pretrained models)
  • From source (with integrated pretrained models)
• Quickstart
• Workflow Overview
  • ✅ The first Prefect workflow: molecular dynamics
  • 🚀 Parallelize benchmarks at scale
  • List of modular tasks
• Contribute and Development
  • Add new MLIP models
    • External ASE Calculator (easy)
    • Hugging Face Model (recommended, difficult)
  • Add new benchmark
• Citation

Announcement

• [Sep 18, 2025] 🎊 MLIP Arena is accepted as a Spotlight (top 3.5%) at NeurIPS! 🎊
• [Apr 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)

pip install mlip-arena

From source (with integrated pretrained models)

Linux

# (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

Mac

# (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

Instruction for individual benchmark is provided in the README in each corresponding folder under /benchmark.

Workflow Overview

✅ The first Prefect workflow: molecular dynamics

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

from mlip_arena.models import MLIPEnum
from mlip_arena.tasks import MD
from mlip_arena.tasks.utils import get_calculator

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=get_calculator(
            model,
            calculator_kwargs=dict(), # passing into calculator
            dispersion=True,
            dispersion_kwargs=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
        total_time=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.

...
from prefect import flow

@flow
def run_all_tasks:

    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 modular 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
• NEB_FROM_ENDPOINTS: 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 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.

Hugging Face Model (recommended, difficult)

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

Benchmark Evaluation Pipeline (Model Submission)

Once your model is ready (either registered or initialized as a custom ASE Calculator), you can run the core benchmark suite on a SLURM cluster:

1. Move into the benchmarks/ directory:

   cd benchmarks
   

2. Open and modify the submit_model.py template script. Under the USER CONFIGURATION section:
   • Provide your MODEL (as a registered string or custom ASE Calculator instance).
   • Adjust the SLURM_CONFIG parameters for your specific HPC allocation (including any conda environments or module loads in the job_script_prologue).
3. Submit the pipeline:

   python submit_model.py
   

   This will dynamically distribute and run the core benchmarks (diatomics, EOS bulk, and E-V scans) via a Dask-Jobqueue on your SLURM cluster.

Add new benchmark

Citation

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

@inproceedings{
    chiang2025mlip,
    title={{MLIP} Arena: Advancing Fairness and Transparency in Machine Learning Interatomic Potentials via an Open, Accessible Benchmark Platform},
    author={Yuan Chiang and Tobias Kreiman and Christine Zhang and Matthew C. Kuner and Elizabeth Jin Weaver and Ishan Amin and Hyunsoo Park and Yunsung Lim and Jihan Kim and Daryl Chrzan and Aron Walsh and Samuel M Blau and Mark Asta and Aditi S. Krishnapriyan},
    booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
    year={2025},
    url={https://openreview.net/forum?id=SAT0KPA5UO}
}