Path / file download issues:

MACE really likes to download data in ~/.cache/mace, which we cannot easily share. I have linked that folder to another folder in scratch, you can do the same by running (assuming the cache doesn't already exist in your home directory)

ln -s /leonardo_scratch/fast/IIT24_AtomSim/franken/franken_cache/mace_cache ~/.cache/mace

so you shouldn't need to download anything new.

Configurations

• water_ft_all.yaml finetune all of MACE
• water_ft_head.yaml finetune the readout head
• water_ft_head_replay.yaml finetune the readout head using multihead finetuning with replays from the MP data

The configs are setup to save results in /leonardo_work/IIT24_AtomSim/franken/rebuttal/mace_finetune

The data is all stored in /leonardo_scratch/fast/IIT24_AtomSim/franken/franken_cache/water:

• train_1024_seed1.xyz contains 1k samples
• train_32_seed1.xyz contains 32 samples
• train_8_seed1.xyz contains 8 samples these have been generated with the create_subsampled_dataset.py script.

Running ft

PYTHONPATH='.' python -m mace.cli.run_train --config mace/configs/water_ft_all.yaml

With slurm (make sure to check the slurm parameters make sense)

./slurm/run_with_cfg.sh mace/configs/water_ft_all.yaml

New scripts

For dataset stuff:

• rewrite_dset_keys.py : rewrites energy to REF_energy, forces to REF_forces in the xyz
• create_subsampled_dataset.py : creates subsampled datasets

Evaluation:

• evaluate_models.py : runs validation for all trained models and saves predictions

Training:

• slurm/run_full_exp.py : runs a list of training experiments iterating through random seed and number of samples