Can we load multiple types of atoms into DeePMD-kit?

[IricsDo]

I have 3 combinations of the same elements: Si, H, and C. Specifically the systems:

Si(56)-C(56)-H(52)

Si(56)-C(56)-H(55)

Si(56)-C(56)-H(58)

How can I load these 3 combinations into one place to train the model? My goal is based on these 3 types of data; I can interpolate different energies of the Si, H, and C systems.

For example, predict the energy of the Si(56)-C(56)-H(100) system.

Now the way I can think of is to read them from my output.out file (for each type with a different number of H) from Siesta software. Then convert them to npy format.
label_52 = dpdata.LabeledSystem("52_raw/data", fmt="deepmd/npy")
label_55= dpdata.LabeledSystem("55_raw/data", fmt="deepmd/npy")
label_58 = dpdata.LabeledSystem("58_raw/data", fmt="deepmd/npy")

Then I will use the NumPy library to read these NPY files and combine them into a single NPY file (box, coord, energy, force, virial). But with type.raw, there is only 1. And I used the type.raw file, which has 52 hydrogen atoms.

def combine_npy_files(
    folder_data: list[str], output_folder: str, verbose: bool = False
) -> bool:

    new_output_folder = os.path.join(output_folder, "set.000")
    create_folder(new_output_folder)
    copy_one_file(
        os.path.join(folder_data[0], "type.raw"),
        os.path.join(output_folder, "type.raw"),
    )
    copy_one_file(
        os.path.join(folder_data[0], "type_map.raw"),
        os.path.join(output_folder, "type_map.raw"),
    )
    LOGGER = ServerLogger()

    npy_files = ["virial.npy", "force.npy", "energy.npy", "coord.npy", "box.npy"]

    for data_type in npy_files:
        data_combine = list()
        for folder in folder_data:
            data_combine.append(np.load(os.path.join(folder, "set.000", data_type)))

        # if all(arr.shape == data_combine[0].shape for arr in data_combine):
        #     combined_data = np.vstack(data_combine)
        #     np.save(os.path.join(new_output_folder, data_type), combined_data)
        
        if data_combine:
            # Temp fix incorrect shape of data
            # Truncate to the smallest number of columns
            min_cols = min(arr.shape[1] for arr in data_combine)
            data_combine_aligned = [arr[:, :min_cols] for arr in data_combine]
            combined_data = np.vstack(data_combine_aligned)
            np.save(os.path.join(new_output_folder, data_type), combined_data)
        else:
            raise ValueError("The list is empty shapes. Cannot be stacked.")


        if verbose:
            LOGGER.log(f"Combined {data_type} and saved to {new_output_folder}")

    return True

But the results were not as expected. The model predicted the correct trend but the wrong value.

Loss:

I am using the code:

label_52_55_58 = dpdata.LabeledSystem("52_55_58_raw/data", fmt="deepmd/npy")

print(label_52_55_58)

and the result:

Another result from code:

for i in range(label_52_55_58.get_nframes()):
  if i % 100 == 0:
    print(f"Frame {i}: {label_52_55_58[i].data['atom_numbs']}")

My desired output in the above two code snippets should be 55 and 58 since my dataset has all 3 types.

How can I load 3 different types of data into the same LabeledSystem?

I hope everyone can help me.

[wanghan-iapcm]

You cannot combine systems with different total number of atoms into one. https://docs.deepmodeling.com/projects/deepmd/en/stable/data/index.html

deepmd-kit supports training with multiple labeled data systems. https://docs.deepmodeling.com/projects/deepmd/en/stable/train/train-input.html#argument:training/training_data/systems

[IricsDo]

Thank you too much. Your solution is correct. I will close and reopen if I get any trouble in the future.