niklastr/langevin_flow_maps

Langevin Flow Maps

Model checkpoints and peptide simulation data accompanying Langevin Flow Maps: Efficient Molecular Dynamics and Transition Path Sampling.

Repository contents

Directory	Contents
checkpoints/	Transferable LFM weights for approximately 9M- and 27M-parameter models, with reconstruction metadata and training state.
openmm/	Amber14/OBC1 comparison trajectories for nine peptides, plus three ANKSMIEA exploration validation runs.
rollouts/	BAOAB baseline and large-step trajectories
msm/reference/	Frozen reference state definitions and MSM artifacts for the nine peptides, with coordinate transforms, lag times and selection provenance.

The peptide systems are AA, DI, HM, FESD, ARIP, WNMA, PGESTAES, GNDLVTVI and ANKSMIEA, spanning lengths of 2, 4 and 8 amino acids.

Model checkpoints

The 9M and 27M model families include completed checkpoints at three training budgets: 100A, 200A and 500A. These labels refer to the number of atom exposures per model parameter during training. The exact exposure for 500A is approximately 481.5 atoms per parameter. Base checkpoints and a longer-horizon 9M base model are also available.

For inference, download model.eqx and meta.json together. model.eqx contains the EMA inference weights; meta.json records the architecture and model configuration. training.eqx contains the full training tensor state.

See the checkpoint catalog for model variants, training steps and exact exposure counts. The historical 9M_480A name in rollout paths refers to the checkpoint released here as 9M_500A.

Trajectories and manuscript comparisons

OpenMM: the production collection contains three seeds for each peptide, using Amber14/OBC1 at 310 K and a 1 fs timestep with an underdampened diffusivity. Directory names such as DI_1us_seed1 indicate the intended production target; the actual saved duration is recorded in each run's metadata. The three ANKSMIEA_exploration_* runs are 100 ns validation simulations initialized from conformations discovered by an LFM.

Learned-force BAOAB: folders named baoab1fs_h0 contain the force-field baseline obtained by querying the LFM at h=0 and integrating with BAOAB at 1 fs. These runs supply the LFM / BAOAB row of the manuscript's peptide sampling table. That comparison uses the first 10 million integration steps, or 10 ns, of each trajectory; the released trajectories extend to approximately 50 ns.

Reference states: the sampling table compares empirical frame and lagged-transition counts in frozen states defined from Many Peptides reference trajectories. Their source provenance and selection diagnostics are recorded in the MSM documentation and artifacts.

The current trajectory release covers the OpenMM comparisons, exploration validation simulations and learned-force BAOAB baseline. The separate corrected LFM forward trajectories used in other manuscript figures are not included in this release.

Using the trajectory data

Each run provides NumPy arrays for positions (nm), velocities (nm/ps) and energies (kJ/mol), together with its simulation settings in config.json. BAOAB rollout potential energies come from the learned model; OpenMM energies come from Amber14/OBC1.

Use snapshots.json for the committed frame count and actual saved duration, and steps.npy for explicit OpenMM timestamps. Keep independent seeds separate. publication.json records export provenance; failure.json, when present, records instability or early termination. Segment and restart histories are consolidated in segments.tar; extract it inside the run directory before analyses that inspect those boundaries.

Citation

Please cite the paper when using these models or data:

@article{mccallum2026langevin,
  title={Langevin Flow Maps: Efficient Molecular Dynamics and Transition Path Sampling},
  author={McCallum, Sam and Rindtorff, Niklas and Tong, Alexander and Foster, James},
  year={2026},
  eprint={2610.05998},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2610.05998}
}