Before you start

This part is for students who needs MACE for their research. If you don't use MACE, then you can ignore this page.
You should finish the orientation before proceed this tutorial. Make sure that you have installed MACE in Python venv.

Fine-tuning a MACE model¶

Why fine-tuning?¶

Foundation model are not accurate enough to solve our problem.

It is known (http://arxiv.org/abs/2405.07105) that foundation model (MACE-MP-0) will underestimate the activation energy barrier. The reason is because the training data are mostly coming from the region near the energy minimum.
If your calculation used different input parameters compared to Materials Project, the energy and forces might not be comparable.

Tips¶

You must supply your DFT computed atomic energies of your system. For example, for \(\ce{Al(HCOO)3}\), you need to supply energies of Al, H, C, O in a large empty box (at least \(10\times 10\times 10 Å^3\)) as --E0s.
You should use --multiheads_finetuning=True for better fine tuning results.
You should use --enable_cueq=True for better performance.

How to fine tune a model¶

You should refer to the official documentation of MACE: https://mace-docs.readthedocs.io/en/latest/guide/finetuning.html. An example of the trainig script is shown below:

mace_run_train \
    --name="name" \
    --multiheads_finetuning=True \
    --foundation_model="mace_agnesi_medium.model"\
    --train_file="data.xyz" \
    --valid_fraction=0.1 \
    --energy_weight=1 \
    --forces_weight=10 \
    --compute_stress=True \
    --stress_weight=1 \
    --stress_key='stress' \
    --swa \
    --swa_forces_weight=100\
    --E0s="{32:-13.928397,3:-0.40914546,8:-0.62290835,15:-4.5251547,16:-5.1361696}"\
    --model="MACE" \
    --scaling='rms_forces_scaling' \
    --lr=0.01 \
    --ema \
    --ema_decay=0.99 \
    --batch_size=24 \
    --max_num_epochs=100 \
    --amsgrad \
    --device="cuda" \
    --seed=1234 \
    --scheduler_patience=5 \
    --patience=50 \
    --enable_cueq=True \
    --default_dtype="float64" > train.log