Bayesic Force Fields

Bayesic Force Fields (BFF) learns fixed-charge force-field parameters for GROMACS by Bayesian inference. It runs classical MD for many sampled parameter sets, compares quantities of interest (QoIs) such as RDFs or free energy profiles with a reference simulation, fits Gaussian-process surrogates to them, and samples the posterior distribution of the parameters.

Workflow

Each step is one command with one YAML config:

  bff build                 build and equilibrate the GROMACS systems
      │
      ▼
  reference MD              outside BFF: AIMD or an MLIP (see below)
      │
      ▼
  bff sample-parameters     draw parameter sets, run classical MD for each
      │
      ▼
  bff build-qoi-datasets    compute the QoIs of every sample and of the reference
      │
      ▼
  bff fit-lgp               fit a Gaussian-process surrogate per QoI
      │
      ▼
  bff learn                 sample the posterior of the parameters (MCMC)
      │
      ▼
  bff validate              rerun MD with posterior parameters

BFF does not generate the reference data. You run the reference simulation yourself; we recommend ab initio MD or a machine-learned interatomic potential (MLIP), ideally a foundation model fine-tuned for your system. For fine-tuning, the repository provides scripts/label_structures.py, which labels MD frames with CP2K on Slurm. It is not installed by pip; download it from the repository. See Reference data.

Install

mamba create -n bfflearn python=3.10 pip
mamba activate bfflearn
pip install torch          # pick the build for your CPU/GPU: https://pytorch.org/get-started/locally/
pip install bfflearn

The MD stages need GROMACS (with Colvars or PLUMED for biased systems). Notebook examples need pip install "bfflearn[notebook]".

Quick start

bff examples           # copy the examples matching your BFF version
cd examples/acetate    # full workflow: aqueous acetate and calcium acetate

Every stage of the example is a directory with its config; run bff <stage> config.yaml inside it. The acetate walkthrough explains each step.

Learned parameters

Names in one label share one value. Charges can be tied by residue- or system-level charge constraints.

Repository

BayesicForceFields/
├── bff/                     the Python package and the `bff` command
│   ├── workflows/           one package per stage
│   ├── qoi/                 QoI routines (rdf, hydrogen_bonds, custom) and datasets
│   ├── bayes/, mcmc/        surrogates, likelihoods, posterior sampling
│   └── domain/, io/         data models and file formats
├── examples/
│   ├── acetate/             full MD workflow with an RDF and a PMF
│   ├── arbitrary-data/      notebook: learn from your own tabular data
│   └── neon-mie-lgpmd/      notebook: learn from published RDFs
├── scripts/
│   └── label_structures.py  CP2K labeling for MLIP fine-tuning (not installed by pip)
├── docs/                    documentation site
└── tests/

Documentation

vojtechkostal.github.io/BayesicForceFields: how BFF works, all settings, examples, and development. See also the changelog, contributing, support, and security.

Citation

Kostal, V.; Shanks, B. L.; Jungwirth, P.; Martinez-Seara, H. Bayesian Learning for Accurate and Robust Biomolecular Force Fields. J. Chem. Theory Comput. 2026, 22 (5), 2652-2663. doi:10.1021/acs.jctc.5c02051

The code used for the paper is archived as v0.0.1.

License

GNU GPL v3