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},
}
```