bfflearn 0.1.0.dev1

Bayesian workflows for fixed-charge molecular force fields

Bayesic Force Fields

Bayesic Force Fields (BFF) is a workflow-oriented Python package for learning fixed-charge molecular force-field parameters from molecular dynamics observables. It combines system preparation, sampled MD campaigns, QoI analysis, surrogate training and posterior inference in one toolchain.

The public CLI is centered around six workflows:

• bff prepare
• bff simulate
• bff qoi
• bff train
• bff learn
• bff validate

Documentation lives under docs/ and is intended to be published with MkDocs on GitHub Pages.

Published documentation: vojtechkostal.github.io/BayesicForceFields

How to Cite

If you use BFF, please cite:

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.
https://doi.org/10.1021/acs.jctc.5c02051

Paper: Bayesian Learning for Accurate and Robust Biomolecular Force Fields

Preprint: arXiv:2511.05398

Installation

Recommended user installation:

mamba create -n bfflearn python=3.10 pip
mamba activate bfflearn

Install a matching PyTorch build for your machine before training or learning. The recommended way is to use the selector on the official PyTorch install page: https://pytorch.org/get-started/locally/

Example for Linux with CUDA 12.6:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126

Then install BFF from PyPI:

pip install bfflearn

External tools are still required for full workflows:

• Gromacs for prepare, simulate, and validate
• CP2K for staged reference calculations
• PLUMED only for PLUMED-biased systems
• PyTorch installed separately for train, learn, and posterior notebooks

PyTorch is not installed by default because the appropriate CPU or CUDA build depends on the target machine. Install the matching PyTorch build first, then install BFF.

For development work on the repository itself, use:

mamba env create -f environment.yaml
mamba activate bfflearn

If you prefer to start from an existing environment instead:

pip install -e ".[dev,docs,notebook]"

Quick Start

The acetate example in examples/acetate/ shows the intended stage order:

cd examples/acetate/01-prepare/colvars
bff prepare config.yaml

cd ../../03-training-trjs
bff simulate config-local.yaml

cd ../04-qoi
bff qoi config.yaml

cd ../05-train-lgp
bff train config.yaml

cd ../06-learn
bff learn config.yaml

Validation is configured separately in stage 08:

cd ../08-validate
bff validate config.yaml

Two notebooks are included in the example:

• 06-learn/interactive.ipynb shows interactive surrogate training, posterior sampling, and posterior sample export.
• 07-visualize/visualize.ipynb focuses on plotting and inspection only.

Repository Layout

• bff/ contains the package code.
• examples/acetate/ contains the worked example.
• data/ contains repository example inputs.
• docs/ contains the documentation source.

Documentation Locally

Preview the docs locally with MkDocs:

mkdocs serve

Build the static site with:

mkdocs build --strict

Shortcuts are also available:

make docs
make docs-build

Shell Completion

When bff runs inside an activated conda environment, it writes a small completion hook for bash and zsh into that environment. After the first bff run, reactivate the environment once:

conda deactivate
conda activate bfflearn

After that, bff <TAB> should offer the public workflow commands.

Development and Release

Packaging, docs, and deployment configuration live in:

• pyproject.toml
• environment.yaml
• .github/workflows/

The release and publication strategy is documented in docs/development.md.

License

BFF is distributed under the GNU GPL v3. See LICENSE.