Inquire about mu_alpha branch progress

[pouilhood]

Hello,

I want to use the MACE model to fitting dielectric properties. How is the development of this feature progressing in the mu_alpha branch? When will it be available? We look forward to hearing from you.

Thank you!

[ilyes319]

Hello @pouilhood, indeed our current development are concentrated on the https://github.com/ACEsuit/mace/tree/mu_alpha branch. It is not documented yet, but you can fit polarizability, dipoles, charges and predict different derivatives of these quantities. We aim to merge it for the release 0.3.8.

[ilyes319]

Here is an example to fit dipole and polarizabilities:

python ./scripts/run_train.py ^
    --name="MACE_dipole_pol" ^
    --train_file="D:\Work\mace_mu_test\train.xyz" ^
    --valid_file="D:\Work\mace_mu_test\valid.xyz" ^
    --test_dir="D:\Work\mace_mu_test\test.xyz" ^
    --model="AtomicDielectricMACE" ^
    --E0s="average" ^
    --num_channels=8 ^
    --max_L=1 ^
    --r_max=5.0 ^
    --loss="dipole" ^
    --dipole_weight=100.0 ^
    --MLP_irreps="16x0e + 16x1o + 16x2e" ^
    --polarizability_weight=100.0 ^
    --dipole_key="REF_mu"  ^
    --polarizability_key="REF_alpha" ^
    --weight_decay=5e-10 ^
    --clip_grad=1.0 ^
    --batch_size=1 ^
    --valid_batch_size=1 ^
    --max_num_epochs=3 ^
    --scheduler_patience=20 ^
    --patience=50 ^
    --eval_interval=1 ^
    --ema ^
    --error_table="DipoleRMSE" ^
    --default_dtype="float64" ^
    --device=cpu ^
    --seed=123 ^
    --save_cpu ^
    --compute_atomic_dipole ^
    --compute_polarizability