--- 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}, } ```