Can I blend ZBL potential during the fine-tuning stage?

[Ruoyan0103]

Hi,

I would like to ask if I can add a tabulated ZBL table during the fine-tuning stage of DPA-3.1-3M.pt pretrained potential?

Thanks

[Ruoyan0103]

Hi,

I just found it mentioned here. The model section in input.json must be the same as that in the pretrained model.

So the answer is no.

[Ruoyan0103]

Thanks. Looking forward to it.

[wanghan-iapcm]

please track #4849

[Ruoyan0103]

please track #4849
Thanks! I tried it. I have a question: should we expect the energy to match exactly with the one from the provided ZBL file?
I use the code like this to test my fine-tuned model:

import numpy as np
calculator = DP("model.ckpt-200.pt") # fine-tuned model
with open(f'ZBL_dp-ft.table', 'w') as f:
    for d in np.arange(0.001, 0.03, 0.001):
        atoms = Atoms('O2', positions=[(0, 0, 0), (d, 0, 0)], cell=[100, 100, 100], pbc=True)
        atoms.calc = calculator
        dp_e1 = atoms.get_potential_energy()

I noticed that, for the same dimer distance, the fine-tuned model gives different energies compared to those in the provided file. The relative difference is around 0.5% compared to the values in the table.

[anyangml]

please track #4849
Thanks! I tried it. I have a question: should we expect the energy to match exactly with the one from the provided ZBL file?
I use the code like this to test my fine-tuned model:

import numpy as np
calculator = DP("model.ckpt-200.pt") # fine-tuned model
with open(f'ZBL_dp-ft.table', 'w') as f:
    for d in np.arange(0.001, 0.03, 0.001):
        atoms = Atoms('O2', positions=[(0, 0, 0), (d, 0, 0)], cell=[100, 100, 100], pbc=True)
        atoms.calc = calculator
        dp_e1 = atoms.get_potential_energy()

I noticed that, for the same dimer distance, the fine-tuned model gives different energies compared to those in the provided file. The relative difference is around 0.5% compared to the values in the table.

Please take a look at the Doc
My intuition is that the model predictions should not match the tabulated value exactly for the following reason:

1. if the pair wise distance is between the lower and upper bound, the model prediction is a linear combination of the tabulated data and the DP model.
2. even if it's below the lower bound, and prediction are dominated by the tabulated data, there is still a energy bias term E0.

[Ruoyan0103]

please track #4849
Thanks! I tried it. I have a question: should we expect the energy to match exactly with the one from the provided ZBL file?
I use the code like this to test my fine-tuned model:

import numpy as np
calculator = DP("model.ckpt-200.pt") # fine-tuned model
with open(f'ZBL_dp-ft.table', 'w') as f:
    for d in np.arange(0.001, 0.03, 0.001):
        atoms = Atoms('O2', positions=[(0, 0, 0), (d, 0, 0)], cell=[100, 100, 100], pbc=True)
        atoms.calc = calculator
        dp_e1 = atoms.get_potential_energy()

I noticed that, for the same dimer distance, the fine-tuned model gives different energies compared to those in the provided file. The relative difference is around 0.5% compared to the values in the table.

Please take a look at the Doc My intuition is that the model predictions should not match the tabulated value exactly for the following reason:

1. if the pair wise distance is between the lower and upper bound, the model prediction is a linear combination of the tabulated data and the DP model.

2. even if it's below the lower bound, and prediction are dominated by the tabulated data, there is still a energy bias term E0.

Thanks! I checked the document and I agree with you.