Speed and accuracy benchmark of foundational machine learning interatomic potentials (MLIPs) for biomolecular simulation, following the protocol of Eastman et al. (2026).
Accuracy — mean absolute error on the SPICE v3 test set (800 structures across small ligands, large ligands, pentapeptides, and dimers).
Speed — NVT MD throughput (ms/step, ns/day) on 10 systems ranging from 22 to 100k atoms, measured on three AWS GPU instances: g7e.4xlarge (RTX PRO 6000 Blackwell), g6e.4xlarge (L40S), and g5.4xlarge (A10G).
| Model family | Models |
|---|---|
| MACE | MACE-OFF23(S/L), MACE-OFF24(M), MACE-MH-1, MACE-OMOL-0, Polar-1(S/M/L) |
| AceFF | AceFF-1.1, AceFF-2.0 |
| FeNNix | FeNNix-Bio1(S/M) |
| UMA | UMA-s-1.2, UMA-m-1.1 |
| Orb | Orb-v3-omol |
| AIMNet2 | AIMNet2 |
| Egret | Egret-1 |
| MACELES | MACELES-OFF |
# Set up conda environments (Python 3.11, PyTorch cu128)
bash benchmark/setup.sh
# Run a model (example)
conda activate ase-mace
python benchmark/run_mace.pyEastman, P., Pretti, E., & Markland, T. E. (2026). Accuracy and Efficiency Benchmarks of Pretrained Machine Learning Potentials for Molecular Simulations. Journal of Chemical Theory and Computation. https://doi.org/10.1021/acs.jctc.6c00130