Vibrational spectra prediction

[nitbha007]

Hello,
I want to use the "EnergyDipolesMACE" MACE model to predict the vibrational spectra. Can you direct me to the script which you used to obtain the spectra in your paper? Are you using auto-correlation function path or taking the derivative of the dipole moment to have the spectra.

Thank you!

[davkovacs]

Hi! Can you say please which spectra you are talking about?

[nitbha007]

Hi, IR and Raman spectra.

[davkovacs]

But you are looking for a script which I used to create a spectra in a paper? What paper and spectra do you mean :D

[nitbha007]

This paper: https://doi.org/10.1063/5.0155322

[davkovacs]

In that paper there is no IR or Raman spectra. I just computed the Fourier-transform of the velocity-velocity autocorrelation function.

If you want IR and Raman I would recommend just using a simple MACE or ScaleShiftMACE to generate the trajectory and use a separate AtomicDipolesMACE to compute the dipoles and polarizabilities after the trajectory is generated.

Please use this PR for trainign the dipole prediction model (I will try to merge it after Christmas ) #149

And here is an example script for fitting the model for the dipoles and polarisabilities:

python /PATH_TO_MACE/mace/scripts/run_train.py \
    --name="MACE_mu_alpha" \
    --train_file="train.xyz" \
    --valid_fraction=0.05 \
    --test_file="test.xyz" \
    --model="AtomicDipolesMACE" \
    --num_channels=32 \
    --max_L=2 \
    --r_max=5.0 \
    --loss="dipole" \
    --dipole_weight=100.0 \
    --polarizability_weight=100.0 \
    --dipole_key="mu" \
    --polarizability_key="alpha" \
    --batch_size=2 \
    --valid_batch_size=2 \
    --max_num_epochs=75 \
    --scheduler_patience=6 \
    --patience=12 \
    --ema \
    --error_table='DipoleRMSE' \
    --default_dtype="float64"\
    --device=cuda \
    --seed=123 \
    --restart_latest \
    --save_cpu \

And to evaluate the trajectory youu can use something like this:

python /PATH_TO_MACE/mace/scripts/eval_mu_alpha.py \
    --configs="md_simulation.xyz" \
    --model="/PATH_TO_MODEL/MACE_mu_alpha.model" \
    --output="mu_alpha_of_MD.xyz" \
    --device=cuda \
    --batch_size=128 \

Please make sure to adjust your batch size depending on the size of the training set for training and the size of the cells / molecules for testing.

[nitbha007]

Thank you. I will use the recommended PR!

[nitbha007]

Hello @davkovacs, Is there a quick way to log the error table at the end of model training? I saw that the logging information is commented out for writing the error about training, validation and test in the run_train.py script in this PR.

Thank you!

[davkovacs]

I probably didn't have time to implement that. Feel free to add it and submit PR.

[mahrossi]

hey @davkovacs - just in case I want to use this branch/PR to predict only the polarizability (no dipoles) do you foresee any issues? I will try hands-on tonight or tomorrow, but just in case you see this and/or have any tips, I would be super grateful.

[mahrossi]

(the loss flattened at some value which gives larg-ish errors though (~12% or less in the individual tensor components) - probably some problem with my input, which I will chase

[ilyes319]

What unit do you currently use? It could have a numerical impact.

[mahrossi]

I am using bohr^3 for the polarizabilities (so, atomic units). any suggestions?

[ilyes319]

Could you try Ang^3? It might help with normalization (not sure).

[mahrossi]

yes I'll try it. I noticed that some scalings and so on are only done on the dipoles (which I am not using here). can that be a problem? sorry, I just read the code quickly, so I might have gotten the wrong impression.