janus-core

Tools for machine learnt interatomic potentials

Features in development

• Support for multiple MLIPs
  • MACE
  • M3GNET
  • CHGNET
• Single point calculations
• Geometry optimisation
• Molecular Dynamics
  • NVE
  • NVT (Langevin(Eijnden/Ciccotti flavour) and Nosé-Hoover (Melchionna flavour))
  • NPT (Nosé-Hoover (Melchiona flavour))
• Nudge Elastic Band
• Phonons
  • Phonopy
• Equation of State
• Training ML potentials
  • MACE
• Fine tunning MLIPs
  • MACE
• Rare events simulations
  • PLUMED

The code relies heavily on ASE, unless something else is mentioned.

Development

1. Install poetry
2. (Optional) Create a virtual environment
3. Install janus-core with dependencies:

git clone https://github.com/stfc/janus-core
cd janus-core
pip install --upgrade pip
poetry install --with pre-commit,dev,docs  # install extra dependencies
pre-commit install  # install pre-commit hooks
pytest -v  # discover and run all tests

Manually updating ASE via https://gitlab.com/ase/ase is strongly recommended, as tags are no longer regularly published. For example:

pip install git+https://gitlab.com/ase/ase.git@b31569210d739bd12c8ad2b6ec0290108e049eea

To prevent poetry downgrading ASE when installing in future, add the commit to pyproject.toml:

poetry add git+https://gitlab.com:ase/ase.git#b31569210d739bd12c8ad2b6ec0290108e049eea

Examples

Single Point Calculations

Perform a single point calcuation (using the MACE-MP "small" force-field):

janus singlepoint --struct tests/data/NaCl.cif --arch mace_mp --calc-kwargs "{'model' : 'small'}"

This will calculate the energy, stress and forces and save this in NaCl-results.xyz, in addition to generating a log file, singlepoint.log, and summary of inputs, singlepoint_summary.yml.

Additional options may be specified. For example:

janus singlepoint --struct tests/data/NaCl.cif --arch mace --calc-kwargs "{'model' : '/path/to/your/ml.model'}" --properties energy --properties forces --log ./example.log --out ./example.xyz

This calculates both forces and energies, defines the MLIP architecture and path to your locally saved model, and changes where the log and results files are saved.

[!NOTE] The MACE calculator currently returns energy, forces and stress together, so in this case the choice of property will not change the output.

By default, all structures in a trajectory file will be read, but specific structures can be selected using --read-kwargs:

janus singlepoint --struct tests/data/benzene-traj.xyz --read-kwargs "{'index': 0}"

For all options, run janus singlepoint --help.

Geometry optimization

Perform geometry optimization (using the MACE-MP "small" force-field):

janus geomopt --struct tests/data/H2O.cif --arch mace_mp --calc-kwargs "{'model' : 'small'}"

This will optimize the atomic positions and save the resulting structure in H2O-opt.xyz, in addition to generating a log file, geomopt.log, and summary of inputs, geomopt_summary.yml.

Additional options may be specified. This shares most options with singlepoint, as well as a few additional options, such as:

janus geomopt --struct tests/data/NaCl.cif --arch mace_mp --calc-kwargs "{'model' : 'small'}" --vectors-only --traj 'NaCl-traj.xyz'

This allows the cell to be optimised, allowing only hydrostatic deformation, and saves the optimization trajector in addition to the final structure and log.

For all options, run janus geomopt --help.

Molecular dynamics

Run an NPT molecular dynamics simulation (using the MACE-MP "small" force-field) at 300K and 1 bar for 1000 steps (1 ps):

janus md --ensemble npt --struct tests/data/NaCl.cif --arch mace_mp --calc-kwargs "{'model' : 'small'}" --temp 300 --steps 1000 --pressure 1.0

This will generate several output files:

• Thermodynamical statistics every 100 steps, written to NaCl-npt-T300.0-p1.0-stats.dat
• The structure trajectory every 100 steps, written to NaCl-npt-T300.0-p1.0-traj.xyz
• The structure to be able to restart the dynamics every 1000 steps, written to NaCl-npt-T300.0-p1.0-res-1000.xyz
• The final structure written to NaCl-npt-T300.0-p1.0-final.xyz
• A log of the processes carried out, written to md.log
• A summary of the inputs and start/end time, written to md_summary.yml.

Additional options may be specified. For example:

janus md --ensemble nvt --struct tests/data/NaCl.cif --steps 1000 --timestep 0.5 --temp 300 --minimize --minimize-every 100 --rescale-velocities --remove-rot --rescale-every 100 --equil-steps 200

This performs an NVT molecular dynamics simulation at 300K for 1000 steps (0.5 ps), including performing geometry optimization, rescaling velocities, and removing rotation, both before beginning dynamics and at steps 100 and 200 of the simulation.

janus md --ensemble nve --struct tests/data/NaCl.cif --steps 200 --temp 300 --traj-start 100 --traj-every 10 --traj-file "example-trajectory.xyz" --stats-every 10 --stats-file "example-statistics.dat"

This performs an NVE molecular dynamics simulation at 300K for 200 steps (0.2 ps), saving the trajectory every 10 steps after the first 100, and the thermodynamical statistics every 10 steps, as well as changing the output file names for both.

For all options, run janus md --help.

Heating

Run an NVT heating simultation from 20K to 300K in steps of 20K, with 10fs at each temperature:

janus md --ensemble nvt --struct tests/data/NaCl.cif --temp-start 20 --temp-end 300 --temp-step 20 --temp-time 10

The produced statistics and trajectory files will indicate the heating range NaCl-nvt-T20.0-T300.0-stats.dat, NaCl-nvt-T20.0-T300.0-traj.xyz. There will also be final structure files at each temperature point:

NaCl-nvt-T20.0-final.xyz
NaCl-nvt-T40.0-final.xyz
...
NaCl-nvt-T300.0-final.xyz

MD can also be carried out after heating using the same options as described in Molecular dynamics. For example:

janus md --ensemble nvt --struct tests/data/NaCl.cif --temp-start 20 --temp-end 300 --temp-step 20 --temp-time 10 --steps 1000 --temp 300

This performs the same initial heating, before running a further 1000 steps (1 ps) at 300K.

When MD is run with heating the trajectory NaCl-nvt-T20.0-T300.0-T300.0-traj.xyz and statistics NaCl-nvt-T20.0-T300.0-T300.0-stats.dat files will indicate the heating range and MD temperature (which may be different). With heating and MD trajectories/statistics within the same files.

Additional settings for geometry optimization, such as enabling optimization of cell vectors by setting hydrostatic_strain = True for the ASE filter, can be set using the --minimize-kwargs option:

janus md --ensemble nvt --struct tests/data/NaCl.cif --temp-start 0 --temp-end 300 --temp-step 10 --temp-time 10 --minimize --minimize-kwargs "{'filter_kwargs': {'hydrostatic_strain' : True}}"

Equation of State

Fit the equation of state for a structure (using the MACE-MP "small" force-field):

janus eos --struct tests/data/NaCl.cif --no-minimize --min-volume 0.9 --max-volume 1.1 --n-volumes 9 --arch mace_mp --calc-kwargs "{'model' : 'small'}"

This will save the energies and volumes for nine lattice constants in NaCl-eos-raw.dat, and the fitted minimum energy, volume, and bulk modulus in NaCl-eos-fit.dat, in addition to generating a log file, eos.log, and summary of inputs, eos_summary.yml.

By default, geometry optimization will be performed on the initial structure, before calculations are performed for the range of lattice constants consistent with minimum and maximum volumes supplied. Optimization at constant volume for all generated structures can also be performed (sharing the same maximum force convergence):

janus eos --struct tests/data/NaCl.cif --minimize-all --fmax 0.0001

For all options, run janus eos --help.

Phonons

Calculate phonons with a 2x2x2 supercell, after geometry optimization (using the MACE-MP "small" force-field):

janus phonons --struct tests/data/NaCl.cif --supercell 2x2x2 --minimize --arch mace_mp --calc-kwargs "{'model' : 'small'}"

This will save the Phonopy parameters, including displacements and force constants, to NaCl-phonopy.yml and NaCl-force_constants.hdf5, in addition to generating a log file, phonons.log, and summary of inputs, phonons_summary.yml.

Additionally, the --band option can be added to calculate the band structure and save the results to NaCl-auto_bands.yml:

janus phonons --struct tests/data/NaCl.cif --supercell 2x2x2 --minimize --arch mace_mp --calc-kwargs "{'model' : 'small'}" --band

If you need eigenvectors and group velocities written, add the --write-full option. This will generate a much larger file, but can be used to visualise phonon modes.

Further calculations, including thermal properties, DOS, and PDOS, can also be calculated (using a 2x3x4 supercell):

janus phonons --struct tests/data/NaCl.cif --supercell 2x3x4 --dos --pdos --thermal --temp-start 0 --temp-end 300 --temp-step 50

This will create additional output files: NaCl-thermal.dat for the thermal properties (heat capacity, entropy, and free energy) between 0K and 300K, NaCl-dos.dat for the DOS, and NaCl-pdos.dat for the PDOS.

For all options, run janus phonons --help.

Using configuration files

Default values for all command line options may be specifed through a Yaml 1.1 formatted configuration file by adding the --config option. If an option is present in both the command line and configuration file, the command line value takes precedence.

For example, with the following configuration file and command:

struct: "NaCl.cif"
properties:
  - "energy"
out: "NaCl-results.xyz"
arch: mace_mp
calc-kwargs:
  model: medium

janus singlepoint --struct KCl.cif --out KCl-results.cif --config config.yml

This will run a singlepoint energy calculation on KCl.cif using the MACE-MP "medium" force-field, saving the results to KCl-results.cif.

[!NOTE] properties must be passed as a Yaml list, as above, not as a string.

Training and fine-tuning MACE models

[!NOTE] This currently requires use of the develop branch of MACE. This can be installed by running poetry add git+https://github.com/ACEsuit/mace.git#develop, followed by poetry install.

MACE models can be trained by passing a configuration file to the MACE CLI:

janus train --mlip-config /path/to/training/config.yml

This will create logs, checkpoints and results folders, as well as saving the trained model, and a compiled version of the model.

Foundational models can also be fine-tuned, by including the foundation_model option in your configuration file, and using --fine-tune option:

janus train --mlip-config /path/to/fine/tuning/config.yml --fine-tune

Calculate MACE descriptors

MACE descriptors can be calculated for structures (using the MACE-MP "small" force-field):

janus descriptors --struct tests/data/NaCl.cif --arch mace_mp --calc-kwargs "{'model' : 'small'}"

This will calculate the mean descriptor for this structure and save this as attached information (descriptors) in NaCl-descriptors.xyz, in addition to generating a log file, descriptors.log, and summary of inputs, descriptors_summary.yml.

The mean descriptor per element can also be calculated, and all descriptors, rather than only the invariant part, can be used when calculating the means:

janus descriptors --struct tests/data/NaCl.cif --no-invariants-only --calc-per-element

This will generate the same output files, but additional labels (Cl_descriptor and Na_descriptor) will be saved in NaCl-descriptors.xyz.

For all options, run janus descriptors --help.

License

BSD 3-Clause License

Funding

Contributors to this project were funded by