Settings Overview

Every stage reads one YAML file:

bff <stage> config.yaml

This page lists all groups of settings; each link leads to the full option table. The conventions below apply to every stage.

bff build

• project, gromacs: output directory, log, GROMACS executable.
• systems[]: per system its ID, topology, coordinate templates, box, MDP files, equilibration and production steps, and an optional Colvars or PLUMED bias.

bff sample-parameters

• bounds, n_samples, seed: which parameters are sampled, in which ranges, and how many sets.
• Parameter labels: the syntax for charges, Lennard-Jones sigma and epsilon, and dihedrals.
• charge_constraints[]: total charges that fix one charge parameter.
• Campaign options: campaign_dir, source, gmx_cmd, job_scheduler, store, cleanup, scratch_dir, max_restarts, resume, overwrite.
• systems[]: which built systems to simulate, and for how many steps.
• slurm: job arrays, sbatch options, setup and teardown lines; see also restarts after the time limit.

bff build-qoi-datasets

• training_samples: the campaign, its systems, analyzed frames, and parallel workers.
• reference: per system the reference files (trajectory, or files such as a PMF).
• routines[]: the QoIs; built-in rdf and hydrogen_bonds or your own function.
• run, output: in-memory analysis and the output directory.

bff fit-lgp

• datasets.<name>: per QoI its dataset, the surrogate mean, and an optional fixed noise.
• fit: model directory, model reuse, committee size, test fraction, and the hyperparameter search.

bff learn

• specs, models.<name>: the surrogates and the tolerance per QoI.
• mcmc: steps, warmup, walkers, prior, convergence, resume, device.
• plots, output: plot sampling and labels, output directory.

bff validate

• posterior or parameters: which parameter sets to simulate.
• All campaign options of sample-parameters.

Conventions

• Unknown keys are errors. A misspelled key is reported, never ignored.
• null means omitted. An option set to null takes its default; a required option set to null is reported as missing.
• Paths are relative to the config file, not to the working directory, and ~ is expanded. Input files must exist when the config is loaded; output directories and logs are created as needed.
• Integers must be whole numbers and numbers must be finite; 1e-3 and 1e5 work although YAML reads them as strings.
• Booleans are true or false, not yes, "false", or 0.
• -1 means unlimited where an option allows it.
• IDs (system_id, routine, dataset, and model names) match [a-z0-9][a-z0-9._-]* and are unique within their list.
• Devices are cpu, cuda, cuda:<index>, or mps.

Errors name the offending key by its full path, for example systems[0].nsteps.prod must be an integer >= 1, got 0. In the option tables, required marks options without a default, and nested keys are written with dots: mcmc.warmup is warmup inside mcmc:.