This folder contains an example of how to finetune a machine learning potential from FAIR Chemistry for "Structure to Relaxed Energy and Forces" (S2EF) on the provided High-Entropy Alloy dataset. This model can then be used as a surrogate DFT calculator to rapidly perform structure relaxations.
To run the contents of this folder, you should follow these installation instructions to set up a conda environment and subsequently install a stable version of the fairchem-core package. Next install cheatools from the main folder.
This example uses Lightning Memory-Mapped Databases (LMDBs) as sources for the training, validation and testing of the model. To create these run dft2lmdbs.py which transforms all images in the DFT trajectories to graphs data objects and saves them to LMDBs.
To avoid training a model from scratch we need a checkpoint file to initilize the pre-trained model. The checkpoint used in this example is the EquiformerV2-153M model trained on the OC20-dataset which can be fetched in the checkpoints folder. After fetching the checkpoint file, you have the option of setting up a wandb profile to monitor the finetuning process (see the config file). The finetuning is initialized by running the finetune.py wrapper script. Be mindful that this should be done on a GPU supported machine. For convenience, an already finetuned checkpoint file can also be fetched in the checkpoints folder.
Run test.py to obtain a parity plot of the test results.
Finally run_relaxations.py showcases writing initial structures to an ASE database and relaxing the structures using the finetuned machine learning potential.