Changelog

The public reproduction snapshot for the published study is archived as v0.0.1. Use that tag for exact reproduction of the paper results. The current workflow release is 0.4.0.

0.4.0 - 2026-08-24

The pipeline now uses semantic system IDs and explicit role-based manifests. Snapshot extraction and CP2K labeling are unified under bff label-snapshots; all molecular and isolated-atom CP2K inputs are supplied explicitly by users; the remaining stages are named sample-parameters, build-qoi-datasets, and fit-lgp. This is a breaking config change; follow the pipeline migration guide.

Build systems now include stable virtual-site-free reference assets, label-snapshots writes a detailed results manifest, and validation can draw samples directly from outputs/posterior.pt. Learning copies the authoritative parameter specification into its fixed outputs/ directory.

QoI analysis now parallelizes complete training samples, processes all systems of a sample sequentially, and treats the reference through the same execution path. Custom routines infer file-based execution from declared inputs rather than a separate loader setting.

Static hydrogen-bond selections now reuse their donor topology and possible labels across frames, substantially reducing reference-analysis time. QoI dataset metadata is also guaranteed to remain acyclic during serialization.

Colvars-enabled sampling and validation jobs now rewrite colvars-configfile relative to the GROMACS working directory, fixing missing bias files in staged local and Slurm campaigns.

The numbered examples follow the new stage contract. The self-contained notebooks use CUDA when available and otherwise fall back to CPU. pytest is now installed through the dev extra instead of as a runtime dependency.

See the repository changelog for the complete list of breaking changes and fixes.

0.3.0 - 2026-06-11

Highlights

• Effective observation counts are now configured during bff learn as explicit counts, independent scalar observations, or tolerance-derived curve features.
• The Gaussian likelihood uses an n_eff-weighted mean squared residual, so posterior width is not determined by arbitrary curve binning.
• bff learn writes qoi-marginals.pdf to show which QoIs support different posterior regions.
• Different QoIs may use different numbers of surrogate-training rows.
• Build templates are optional when a system does not need them, and the CLI uses the single bff <command> entry point.
• User-defined analysis routines now work with multiprocessing workers.
• pytest is included in the standard installation.

Warning

The learn configuration and .lgp model format changed in 0.3.0. Refit models created by earlier releases and migrate path-only model entries to the nested format documented in the learn configuration.

0.2.1 - 2026-06-01

Fixed

• Restored Python 3.10 compatibility by replacing the Python 3.11-only typing.Self annotation in Gaussian-process model loading.

0.2.0 - 2026-06-01

Reference Points

The code history contains two useful paper-era comparison points:

• Earlier branches used ParmEd to parse and modify GROMACS topologies.
• The later public v0.0.1 snapshot had already moved to an intermediate gmxtop parser together with MDAnalysis.
• The current line uses gmxtopology and MDAnalysis selections.

Publication Snippet

Relative to the paper-era implementation, Bayesic Force Fields has been refactored into a staged, configuration-driven workflow for system preparation, reference-data generation, force-field sampling, surrogate fitting, posterior learning, and validation. The refactor replaces the external emcee sampling backend with an in-package Torch MCMC implementation, moves topology handling away from the earlier ParmEd-based path, and adds reusable quantity-of-interest datasets, local and Slurm execution, hierarchical charge constraints, broader GROMACS topology updates, and notebook-first examples for externally generated data.

Architecture Changes

Area	Paper-Era Implementation	0.2.0
Posterior sampling	emcee.EnsembleSampler and its HDF backend	Torch parallel Metropolis-Hastings sampler with adaptive proposals
Diagnostics	Autocorrelation handling through emcee results	Checkpoints, restart support, split R-hat, autocorrelation time, and effective sample size
Topologies	ParmEd in earlier branches; intermediate gmxtop in v0.0.1	gmxtopology with MDAnalysis selections
Constraints	One molecular total charge and one implicit charge parameter	Hierarchical residue- or system-level constraints with compatibility checks
Data model	Monolithic training arrays and YAML sidecars	Reusable serialized QoIDataset objects
API	Broad workflow commands	Focused stages from build through validate

Highlights

• Removed the runtime emcee dependency in favor of the Torch MCMC stack.
• Replaced the earlier ParmEd topology path with GROMACS-native handling.
• Split monolithic structures and inference code into focused modules.
• Reconstructable specs.yaml files and hierarchical charge constraints.
• CP2K snapshot collection, bias inputs, and local or Slurm campaigns.
• Reusable quantity-of-interest datasets and notebook-first examples.
• More stable Gaussian-process fitting, posterior learning, and plotting.
• Function-9 GROMACS dihedral updates using labels such as dihedraltype9_3_180.

The repository CHANGELOG.md keeps the complete grouped history.