Changelog#

Unreleased#

Rename ParameterisationSettings to ParamSettings and the parameterisation_settings field to param_settings throughout. This avoids confusion between British/American spellings. YAML configs must update the key from parameterisation_settings: to param_settings:. In #62.
Replace flat ML potential settings with nested mlp_settings in sampling and MSM settings in #58. Legacy flat fields such as ml_potential and ml_potential_kwargs are no longer accepted.
Remove AvailableModels and validate_model_charge_compatibility. presto now accepts any OpenMM-ML model identifier supported by the local OpenMM-ML install and trusts the user to pick a model compatible with their system. Also in #58.
Add runtime-object placeholder handling for mlp_settings.ml_system_kwargs: non-serializable values (e.g ASE Calculator objects) are written as placeholders and loading fails with InvalidSettingsError until placeholders are replaced (for example with from_yaml(..., overwrite=...)) before validation. Again, in #58.

Rewrite documentation with Diátaxis structure: Get started, How-to, Concepts, and Reference sections.
Add new pages: installation guide, quickstart, bespoke SMIRNOFF primer, sampling protocols, type generation, MLP guide, output layout, troubleshooting, and glossary. Add Python API walk-through notebook. - Polish Pydantic Field descriptions in settings to improve auto-generated API reference. Add example invocations to CLI subcommand docstrings. Add Python API examples for running fitting via get_bespoke_force_field(...), including optional OpenMM-ML ASE calculator usage through ml_system_kwargs. In #61.
Clarify charge behavior: automatic molecular charge propagation applies to non-ASE MLPs; for ml_potential="ase" charge must be passed explicitly in ml_system_kwargs.

Update the default env to use Python 3.12 to avoid dm-tree (an Orb dep) build issues. In #60.

Guard against lack of any constaints in the input force field in #56. This allows use of the unconstrained version of Parsely.
Set default model to AIMNet2 in light of current AceFF 2.0 OpenMM-ML issues in #49. Also add README warning about AceFF 2.0.

Make molecule loading more modular and accept SDFs in #55. Note this is a breaking change -- the old param_settings.smiles field must now be replaced by param_settings.molecules.
Add CLI sub-command which shows version in #54.
Add Orb-v3 OMOL model in #49.

Raise an error if too few conformers survive filtering in #38. Fixes #30.
Remove stereochemical information from generated types in #36. This avoids a niche issue where mixing the RDKit and OpenEye toolkits would result in failed type generation for e.g. chiral sulfoxides.

Make linting more strict and improve docstrings in #39
Update aromaticity model used for selecting rotatable torsions for metadynamics to give more intuitive results (ditch MDL and go for RDKit default) in #33.
Improve consistency of how the device arguments are passed around (literal or torch device). #36
Refactor simulation creation logic to reduce duplication and always set the platform to CPU for ML systems (required so that MACE models work with PythonForce). See this commit.
Update environments to OpenMM 8.5 (with PythonForce) and OpenMM-ML. This simplifies the environments required (as we can drop NNPOPs) and means we have have all MLPs in one env. The only issue is that we can now only support CUDA 12.9. #34.

Fixes typo in default settings which meant that linear torsions matching [*:1]~[*:2]=[#6,#7,#16,#15;X2:3]=[*:4] were not excluded from training (#32)

Use MDTraj to calculate torsions rather than doing this manually
Make metadynamics more aggressive, allow user to specify torsions targeted by metadynamics, and include in-ring aliphatic torsions by default in #26

Improve cleanliness/prettiness of CLI output
Reduce memory use during congeneric series fitting in #17
Improve test quality, speed, and coverage in #19
Update environments with explicit python and cuda versions and document installation with different cuda versions
Plot sampling of rotatable torsions in #23

Renamed bespokefit_smee -> presto (Parameter Refinement Engine for Smirnoff Training / Optimisation)
Added more documentation on the method, recommended settings, and outputs.