File size: 2,501 Bytes
e5b8053 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 | ---
license: mit
tags:
- molecular-dynamics
- normalizing-flows
- jax
- peptides
- kcmaps
---
# Kolmogorov-Chapman Maps: checkpoints and data
Checkpoints, reference MD and generated trajectories for
[Kolmogorov-Chapman Maps: Trajectory-Free Learning for Simulating Molecular Dynamics](https://arxiv.org/abs/2611.12880).
Code: [github.com/majhas/kc-maps](https://github.com/majhas/kc-maps). Released under the MIT License.
```bash
hf download majhas/kc-maps --local-dir kc-maps-hf
export KCMAPS_DATA=$PWD/kc-maps-hf/data
python -m kcmaps.simulate checkpoint=kc-maps-hf/checkpoints/al2
python -m kcmaps.evaluate simulation=kc-maps-hf/data/trajectories/kcmaps/AL2_tau30_n16
```
## Contents
| Path | Contents |
| --- | --- |
| `checkpoints/al2`, `al4`, `al6` | KC-maps checkpoints for alanine dipeptide, tetrapeptide and hexapeptide |
| `data/reference/` | reference MD features (energy, backbone torsions, TICA coordinates) |
| `data/trajectories/` | the generated trajectories evaluated in the paper |
| `data/splits/` | training split indices and links to the source datasets |
The reference MD was run with OpenMM `LangevinMiddleIntegrator` at a 1 fs step and saved every 1 ps: 33 µs for AL2,
33.5 µs for AL4 and 32 µs for AL6.
## Generated trajectories
Each run used one GPU-hour on one L40S with 4 CPU cores, compilation and initialization excluded. All runs of a
molecule start from the same start configuration, with momenta drawn independently at the target temperature.
Frames are saved every 1 ps for all methods.
| Run | Method | Settings |
| --- | --- | --- |
| `kcmaps/AL2_tau30_n16` | KC-maps | τ = 30 fs, 16 chains |
| `kcmaps/AL4_tau40_n8` | KC-maps | τ = 40 fs, 8 chains |
| `kcmaps/AL6_tau30_n16` | KC-maps | τ = 30 fs, 16 chains |
| `hfm/AL2_tau7_n16` | HFM | τ = 7 fs, 16 chains, released HFM checkpoint |
| `hfm/AL4_tau7_n64` | HFM | τ = 7 fs, 64 chains, FLOP-matched HFM checkpoint |
| `hfm/AL6_tau5_n1` | HFM | τ = 5 fs, 1 chain, FLOP-matched HFM checkpoint |
| `md/AL2_n16`, `md/AL4_n16`, `md/AL6_n16` | MD | OpenMM `LangevinMiddleIntegrator`, 1 fs, 16 chains |
## Citation
```bibtex
@misc{hassan2026kcmaps,
title={Kolmogorov-Chapman Maps: Trajectory-Free Learning for Simulating Molecular Dynamics},
author={Majdi Hassan and Marta Skreta and Dominique Beaini and Kirill Neklyudov},
year={2026},
eprint={2611.12880},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2611.12880},
}
```
|