UniFF-MD: Universal Machine Learning Force Fields Molecular Dynamics

We present a comprehensive evaluation of six state-of-the-art UMLFFs (CHGNet, M3GNet, MACE, MatterSim, SevenNet, Orb) on a carefully curated dataset, namely MinX, comprising ∼1,500 minerals with experimentally obtained crystal structures and elastic properties. Our analysis is divided into three parts:

• A systematic comparison of model prediction across the minerals dataset.
• A quantitative assessment of temporal evolution during MD simulations.
• Evaluation of elastic constants prediction to study their efficacy of modelling mechanical properties

Installation

Clone this repository and its submodules:

## Ensure CUDA >= 12.1 and GCC Compiler > 9.2 is available 
git clone https://github.com/M3RG-IITD/UniFFBench.git
cd UniFFBench
git submodule update --init --recursive
conda create -n uniff python=3.11.11
conda activate uniff
pip install -r requirments.txt
pip install mace-torch==0.3.10
cd matsciml ## change directory to matsciml folder
pip install --no-deps -e .

• Docker: We provide a Dockerfile inside the docker that can be run to install a container using standard docker commands.
• mamba: We have included a mamba specification that provides a complete out-of-the-box installation. Run mamba env create -n matsciml --file conda.yml, and will install all dependencies and matsciml as an editable install.
• pip: In this case, we assume you are bringing your own virtual environment. Depending on what hardware platform you have, you can copy-paste the following commands; because the absolute mess that is modern Python packaging, these commands include the URLs for binary distributions of PyG and DGL graph backends.

Quickstart

MD Simulation

To run the MD simulation on the test data provided in the repository, first activate your created conda environment and go to the UniFFBench directory and run the following command in .

## Ensure internet is on to load the model checkpoint 
python  md_simulation/experiment_runner.py \
	--model_name mace_pyg \
	--input_dir test_data \
	--index 1 \

Elastic Tensor

To run the elastic tensor on the test data provided in the repository, first activate your created conda environment and go to the UniFFBench directory and run the following command in .

## Ensure internet is on to load the model checkpoint 
python  md_simulation/elastic_tensor_runner.py \
	--model_name mace_pyg \
	--input_dir test_data \
	--index 1 \

Note: If you want to run the simulation on your own data, you must prepare it in the same format as the files in the test_data folder and replace the data path from your own data.

Benchmark Data and CIF Details

The benchmark test data for UniFF-MD are available in the test_data folder and full data can be accessed from Zenodo. Complete details of minerals name, formula and reference are provided in the folder named as CIF_metadata Navigate to the following to access them:

• [test_data Folder]: Contains sample data files used for benchmarking.
• [CIF_metadata Folder]: Contains details of minerals metadata used for benchmarking.

Post processing

once you ran the simulation you can use following script for post processing. Note: Before running the script, make sure to update the root_folder path in the Python file to match the directory of the model you are analyzing.

python  post_processing/trail_post_process.py \

Citations

The MinX dataset used to benchmark the UMLFFs is available at Zenodo. Please cite the following paper if you use the dataset and UniFFBench Framework for your work.

@article{mannan2025evaluatinguniversalmachinelearning,
  title={Evaluating Universal Machine Learning Force Fields Against Experimental Measurements},
  author={Sajid Mannan and Vaibhav Bihani and Carmelo Gonzales and Kin Long Kelvin Lee and Nitya Nand Gosvami and Sayan Ranu and Santiago Miret and N. M. Anoop Krishnan},
  year={2025},
  eprint={2508.05762},
  archivePrefix={arXiv},
  primaryClass={cond-mat.mtrl-sci}
}