MLIP command line workflows for phonons, NEB/MEP, ShakeNBreak giving vasp-improvable results (for computational efficiency) for 5 model families are supported by this package. Furthermore, mlip fine-tuning is supported fully for MACE and ORB model families, with petmad coming soon. The former two support LoRA and REPLAY fine tuning.
mlip-phonons --inputs <config-dir>for phonons, DOS, and band structuresmlip-neb --inputs <config-dir>for NEB / MEP runsmlip-snb --inputs <config-dir>for ShakeNBreak defect workflowsmlip-coup --inputs <config-dir>for phonon-coupling analysis
Examples of SnB, NEB, and Phonon workflows are included.
Three families are included in fine-tuning. LoRA and REPLAY style fine tuning are available for MACE and ORB. Pet-mad is still running into some issues, I am yet to push that work.
All workflows are available to be benchmarked. That is, you can download any number of supported models (list found in SUPPORTED_MODELS.yml, but, generally, any model in the MACE, Orb, PET-MAD, Tensornet, mattersim and CHGNet model families).
If defaults.model_name is a list, the ordinary workflow command fans out once per model and runs each case in the model-specific environment from SUPPORTED_MODELS.yml. This enables you to take any workflow defined by its config.yml file, and then simply append model names to run the same workflow across however many models you like.
For the fine-tuning demos, step 4 shows this usage, with:
mlip-neb --inputs demo/fine_tuning/<family>/4_benchmark --report-benchmark
- Put model files under
assets/models/<family>/ - Register model environments in
SUPPORTED_MODELS.yml
For all model families you are using, install and set up their corresponding environments according to environments.md.
Once you do that, you are done and free to run these workflows. More details on running the workflows can be found here.
Dependent on CUDA.