Skip to content

Trails-MD

Trails-MD is a modular framework for adaptive molecular dynamics campaigns. It runs many short MD walkers, projects saved frames into a fixed or machine-learned collective-variable (CV) space, restarts walkers from informative regions, and repeats the cycle.

Key features

  • Engine-agnostic walkers. OpenMM, GROMACS, and Amber share the same adaptive loop.
  • Fixed or learned sampling spaces. User-defined physical CVs, PCA, TICA, TVAE, Deep-TICA, VAMPnet, SPIB, and the supervised Deep-LDA coordinates — all swappable with one keyword. See Collective variables.
  • VAMP-2 feature selection. Reduce thousands of candidate distances to the few dozen that actually carry the slow dynamics. See Feature selection.
  • Six interchangeable spawning policies. Density, Voronoi, local-outlier-factor, farthest-point, weight-conserving weighted ensemble, and MSM-guided (least-counts × slow-mode leverage × uncertainty). See Concepts.
  • Landscape-adaptive binning. Uniform, density-gradient, minimal adaptive binning (MAB), and MSM-eigenvector schemes. See Adaptive binning.
  • In-loop Markov State Models. Estimate an MSM inside the adaptive loop and stop the campaign on kinetic convergence — implied timescales, VAMP-2, stationary distribution, transition matrix, or Bayesian statistical error — rather than on bin occupancy alone. See MSM & kinetic seeding.
  • Lineage-aware exploration. Every spawned frame stores its parent-child ancestry, so connected transition pathways can be reconstructed from otherwise disjoint exploration stages.
  • Restartable campaigns. Per-iteration checkpoints capture the adaptive model, feature history, sampling state, and walker coordinates; deterministic seeding means a resumed run reproduces an uninterrupted one.
  • HPC scalability. Run on a multi-GPU workstation or dispatch walkers as SLURM / PBS array jobs (execution.backend), with submit-retry against queue limits, configurable GPU request directives, and per-walker GPU isolation checks. See Execution.

The adaptive loop

  run short MD walkers (local / SLURM / PBS)
              |
  extract features / project to CV space
              |
  train or update the CV if space_mode is a learned
  method (PCA / TICA / TVAE / Deep-TICA)
              |
  spawn new walkers (density / Voronoi / LOF / FPS / WE / MSM)
              |
  iteration budget reached, or bin occupancy
  plateaued? --- yes ---> stop
              |
              no
              |
        (back to the top)

Sampling continues until either the configured iteration budget is reached or grid/Voronoi bin occupancy plateaus (see Concepts). After a campaign, representative structures can seed longer unbiased production runs for post-hoc Markov State Model (MSM) construction — see MSM & kinetic seeding.

Where to go next