Trainset Configuration

Source code:

• bff/workflows/configs.py
• bff/workflows/trainset.py
• bff/workflows/runsims.py

Purpose

bff trainset samples parameter vectors, writes a trainset directory, and runs the corresponding GROMACS training campaign.

Minimal Example

mol_resname: ACE
trainset_dir: ./trainset
systems:
  - assets: ../01-prepare/colvars/ace-colvars/training/system-000
    n_steps: 1000
  - assets: ../01-prepare/colvars/ace-colvars/training/system-001
    n_steps: 1000
bounds:
  charge C2: [0.0, 1.0]
  charge O1 O2: [-0.8, -0.3]
implicit_atoms: [C2]
total_charge: -0.8
n_samples: 10
gmx_cmd: gmx
job_scheduler: local

Top-Level Keys

• mol_resname Residue name of the parameterized molecule in the GROMACS topology.
• trainset_dir Output directory for sampled topologies, metadata, and trajectories.
• systems Non-empty list of prepared asset directories plus system-specific MD lengths.
• bounds Mapping from parameter label to lower and upper bounds.
• implicit_atoms Atom group whose charge is determined by the total-charge constraint.
• total_charge Target total molecular charge used when building specs.yaml.
• n_samples Number of force-field vectors sampled for the campaign.
• gmx_cmd GROMACS executable.
• job_scheduler Either local or slurm.
• dispatch If true, launch jobs immediately after staging them.
• compress If true, compress finished simulation outputs.
• cleanup If true, remove temporary files after successful runs.
• store Which trajectory outputs to keep. Defaults to ["xtc"].
• slurm Slurm-only runtime settings.

systems[] Keys

• assets Directory created by bff prepare, for example training/system-000.
• n_steps Production MD length for this prepared system within the sampled campaign.

slurm Keys

• slurm.max_parallel_jobs Maximum number of submitted jobs. -1 means no client-side limit.
• slurm.sbatch Mapping passed to the submission script generator.
• slurm.setup Shell commands executed before the BFF MD job.
• slurm.teardown Shell commands executed after the BFF MD job.

Outputs

bff trainset writes:

• trainset/specs.yaml
• trainset/samples.yaml
• per-sample job configs
• per-sample modified topologies
• stored trajectories and energy files