Quick start


To reduce runtime, this “Quick start” guide uses a fast semiempirical model, “GFN2-xTB”, to generate training data, rather than the “default” method used to train mainline OpenFF force fields.

BespokeFit aims to provide an automated pipeline that ingests a general molecular force field and a set of molecules of interest, and produce a new bespoke force field that has been augmented with highly specific force field parameters trained to accurately capture the important features and phenomenology of the input set.

Such features may include generating bespoke torsion parameters that have been trained to capture as closely as possible the torsion profiles of the rotatable bonds in the target molecule which have a large impact on conformational preferences.

The recommended way to install openff-bespokefit is via the conda package manager. There are several optional dependencies, and a good starting environment is:

conda create -n bespokefit -y -c conda-forge mamba python=3.9
conda activate bespokefit 
mamba install -y -c conda-forge openff-bespokefit xtb-python ambertools 

although several other methods are available.

There are two main routes for creating a bespoke force field using BespokeFit: using the command-line interface or using the Python API.

Using the CLI

The fastest way to start producing a bespoke force field for your molecule of interest is through the command-line interface. A full list of the available commands, as well as help about each, can be viewed by running:

openff-bespoke executor --help 

Of particular interest are the run, launch, submit, retrieve and watch commands.

One-off fits

The run command is most useful if you are wanting to perform a quick one-off bespoke fit for a single molecule using a temporary bespoke executor.


You should only have one run command running at once. If you want to compute bespoke parameters for multiple molecules at once see the section on production fits.

It will accept either a SMILES pattern

openff-bespoke executor run --smiles             "CC(=O)NC1=CC=C(C=C1)O" \
                            --workflow           "default"               \
                            --output             "acetaminophen.json"    \
                            --output-force-field "acetaminophen.offxml"  \
                            --n-qc-compute-workers 2                     \
                            --qc-compute-n-cores   1                     \
                            --default-qc-spec xtb gfn2xtb none

or the path to an SDF (or similar) file

openff-bespoke executor run --file               "acetaminophen.sdf"    \
                            --workflow           "default"              \
                            --output             "acetaminophen.json"   \
                            --output-force-field "acetaminophen.offxml" \
                            --n-qc-compute-workers 2                    \
                            --qc-compute-n-cores   1                    \
                            --default-qc-spec xtb gfn2xtb none

in addition to arguments defining how the bespoke fit should be performed and parallelized.


Sometimes bespoke commands will raise RuntimeError: The gateway could not be reached. This can usually be resolved by rerunning the command a few times.

Here we have specified that we wish to start the fit from the general OpenFF 2.0.0 (Sage) force field, augmenting it with bespoke parameters generated according to the default built-in workflow using GFN2-xTB reference data.


Other available workflow can be viewed by running openff-bespoke executor run --help, or alternatively, the path to a saved workflow factory can also be provided using the --workflow-file flag.

Alternatively, certain options defined by the workflow can be overridden from the CLI. For example, the default specification to use for any new QC calculations can be specified using the --default-qc-spec flag, e.g. --default-qc-spec xtb gfn2xtb none. See the --help for other available overrides.

By default, BespokeFit will use create a single worker for each step in the fitting workflow (i.e. one for fragmenting larger molecules, one for generating any needed reference QC data, and one for doing the final bespoke fit), however extra workers can easily be requested to speed things up:

openff-bespoke executor run --file                 "acetaminophen.sdf" \
                            --workflow             "default"           \
                            --n-fragmenter-workers 2                   \
                            --n-optimizer-workers  2                   \
                            --n-qc-compute-workers 2                   \
                            --qc-compute-n-cores   1                   \
                            --default-qc-spec xtb gfn2xtb none

See the chapter on the bespoke executor for more information about parallelising fits.

Production fits

If you are intending to create bespoke parameters for multiple molecules such as a particular lead series, it is recommended to instead launch a dedicated bespoke executor. This has the added benefits of being able to re-use data from previous fits, such as common QC calculations, and easily retrieve previous bespoke fits.

The first step is to launch a bespoke executor. The executor is the workhorse of BespokeFit, and seamlessly coordinates every step of the fitting workflow from molecule fragmentation to QC data generation:

openff-bespoke executor launch --n-fragmenter-workers 1 \
                               --n-optimizer-workers  2 \
                               --n-qc-compute-workers 4 \
                               --qc-compute-n-cores   1

The number of workers dedicated to each bespoke fitting stage can be tweaked here. In general, we recommend devoting most of your compute power to the QC compute stage as this stage is both the most expensive, and most the parallelisable. See the chapter on the bespoke executor for more information about parallelising fits.

Once the executor has been launched, we can submit molecules to have bespoke parameters trained by the executor using the submit command either in the form of a SMILES pattern:

openff-bespoke executor submit --smiles      "CC(=O)NC1=CC=C(C=C1)O" \
                               --workflow    "default"               \
                               --default-qc-spec xtb gfn2xtb none

or loading the molecule from an SDF (or similar) file:

openff-bespoke executor submit --file        "acetaminophen.sdf"   \
                               --workflow    "default"             \
                               --default-qc-spec xtb gfn2xtb none

The submit command will also accept a combination of the two input forms as well as multiple occurrences of either. After successfully submitting the molecules a table will be printed which maps the unique ID that has been assigned by the executor to the submission to the molecule smiles and input file if appropriate. To keep track of submissions we can also have the table saved to csv by add the corresponding --save-submission-info flag to the command.

The ID’s can be used to check on state of the submission:

openff-bespoke executor watch --id "1"

A full list of submissions currently being processes can be printed with the list command:

openff-bespoke executor list

or if you would only like to inspect those that have failed for example:

openff-bespoke executor list --status errored

Once finished, the final force field can be retrieved using the retrieve command:

openff-bespoke executor retrieve --id          "1"                  \
                                 --output      "acetaminophen.json" \
                                 --force-field "acetaminophen.offxml"

See the results chapter for more details on retrieving the results of a bespoke fit.

Using the API

For the more Python oriented user, or for users who are looking for more control over how the bespoke fit will be performed, BespokeFit exposes a full Python API.

At the heart of the fitting pipeline is the BespokeWorkflowFactory. The BespokeWorkflowFactory encodes the full ensemble of settings that will feed into and control the bespoke fitting pipeline for any input molecule, and is used to create the workflows that fully describe how bespoke parameters will be generated for a specific molecule:

from openff.bespokefit.workflows import BespokeWorkflowFactory
from openff.qcsubmit.common_structures import QCSpec

factory = BespokeWorkflowFactory(
    # Define the starting force field that will be augmented with bespoke 
    # parameters.
    # Change the level of theory that the reference QC data is generated at

Similar to the previous steps, here we override the default “default” QC specification to use GFN2-xTB. If we had Psi4 installed, we could remove the default_qc_specs argument and the factory would instead use our mainline fitting QC method. The default factory will produce workflows that augment the OpenFF 2.0.0 force field with bespoke torsion parameters for all non-terminal rotatable bonds in the molecule that have been trained to quantum chemical torsion scan data generated for said molecule.


See the configuration section for more info on customising the workflow factory.

The workflow factory will ingest any molecule that can be represented by the OpenFF Toolkit’s Molecule class and produce a BespokeOptimizationSchema schema:

from openff.toolkit.topology import Molecule

input_molecule = Molecule.from_smiles("C(C(=O)O)N")  # Glycine

workflow_schema = factory.optimization_schema_from_molecule(

This schema encodes the full workflow that will produce the bespoke parameters for this specific molecule, including details about how any reference QC data should be generated and at what level of theory, the types of bespoke parameters to generate and hyperparameters about how they should be trained, and the sequence of fitting steps (e.g. fit a charge model, then re-fit the torsion and valence parameters using the new charge model) that should be performed.

Such a schema is fed into a BespokeExecutor that will run the full workflow:

from openff.bespokefit.executor import BespokeExecutor, BespokeWorkerConfig, wait_until_complete

with BespokeExecutor(
    n_fragmenter_workers = 1,
    n_optimizer_workers = 1,
    n_qc_compute_workers = 2,
) as executor:
    # Submit our workflow to the executor
    task_id = executor.submit(input_schema=workflow_schema)
    # Wait until the executor is done
    output = wait_until_complete(task_id)

if output.status == "success":
    # Save the resulting force field to an OFFXML file
elif output.status == "errored":
    # OR the print the error message if unsuccessful

The BespokeExecutor not only takes care of calling out to any external programs in your workflow such as when generating reference QC data, it also manages spreading a queue of tasks over a pool of worker threads so that fitting can be executed efficiently in parallel. The BespokeExecutor is described in more detail in its own chapter.

Configuring the workflow factory

There workflow factory is largely customisable in order to accommodate different fitting experiments or protocols that you may wish to use:

from openff.qcsubmit.common_structures import QCSpec

from openff.bespokefit.schema.optimizers import ForceBalanceSchema
from openff.bespokefit.schema.smirnoff import ProperTorsionHyperparameters
from openff.bespokefit.schema.targets import TorsionProfileTargetSchema

factory = BespokeWorkflowFactory(
    # Define the starting force field that will be augmented with bespoke 
    # parameters.
    # Select the underlying optimization engine.
        max_iterations=50, penalty_type="L1"
    # Define the types of bespoke parameter to generate and hyper-parameters 
    # that control how they will be fit, as well as the target reference data 
    # that should be used in the fit.
    # Change the level of theory that the reference QC data is generated at

Once the factory is configured, it can be saved

factory.to_file("workflow-factory.yaml") # or .json

and loaded from disk easily

factory = BespokeWorkflowFactory.from_file("workflow-factory.yaml")

This makes it simple to record and share complex configurations. OpenFF recommends making this file available when publishing data generated using the outputs of BespokeFit for reproducibility. Factories that have been saved to disk can also be used via BespokeFit’s command-line interface.

Check the API docs for full descriptions of the factory’s configurable options.