Graph Machine Learning
flashmd / README.md
frostedoyster's picture
Update with PET-OMATPES model details
606a8e6 verified
---
license: apache-2.0
pipeline_tag: graph-ml
---
# Model Card for Model ID
This repository contains the FlashMD universal models. These are models designed to run accelerated molecular dynamics for chemical systems across the periodic table.
## Model Details
These models offer acceleration with respect to traditional molecular dynamics by allowing the use of larger steps. Each FlashMD model follows the following naming
convention: `flashmd_{mlip}_{timestep}fs.ckpt`, where "mlip" is the name of the MLIP (machine-learned interatomic potential) that was used to generate the training
trajectories for the FlashMD model, and {timestep} is the FlashMD timestep in fs. The corresponding MLIP is also made available as `mlip_{mlip}.ckpt`.
At the moment, we have:
- mlip = "pet-omatpes" (with timestep={1,2,4,8,16,32,64,128}): these are FlashMD models trained on the PET-OMATPES MLIP, and therefore allowing to run FlashMD at the
r2SCAN level of theory
- earlier models used in the reference paper ("flashmd_{timestep}fs.pt", with timestep={1,4,8,16,32,64}), only supported by `flashmd<=0.1.2` and useful to reproduce
older results. These are trained on the PET-MAD MLIP and therefore reproduce molecular dynamics at the PBEsol level of theory
### Model Sources
- **Repository:** https://github.com/lab-cosmo/flashmd
- **Paper:** https://arxiv.org/abs/2505.19350
- **Demo:** https://atomistic-cookbook.org/examples/flashmd/flashmd-demo.html
## How to use
These models are supposed to be used with the flashmd package, which is available on PyPI (`pip install flashmd`).