ASE vs DeepMD inconsistency

[podgaetsky]

I am trying to train a model on an organic compound and apply it in LAMMPS.
When I test the model using the ASE package (ase.Atoms with the DP calculator) I get a poor fit, and the energies are shifted:

When I apply the model in LAMMPS I get that the cell expands without a limit, I assume it happens because the forces aren't computed correctly.

Could you assist me please?

[wanghan-iapcm]

please provide the way of reproducing your observation.

[podgaetsky]

Attached all the files required for reproducing the problem.
DeepMDTrain.zip

[wanghan-iapcm]

That's not enough. please include the version of the code and all the commands you executed (one follows your commands and he/she gets the same ase outputs as yours. )

[podgaetsky]

run the following function in python with correct input:

def plot_e_f_deviation(data_path, model_path=None, model_file='graph_compressed.pb', is_ase=True, tutorial=False):
    if is_ase:
        model = DeepPot(model_path + '/' + model_file)
    else:
        model = DeepPotential(model_path + '/' + model_file)
    all_systems = expand_sys_str(data_path)
    E_pred, F_pred, E_true, F_true = [], [], [], []
    for system in all_systems:
        data = dp.LabeledSystem(system, fmt='deepmd/npy')
        if is_ase:
            energy = np.zeros(data.get_nframes())
            force = np.zeros((data.get_nframes(), len(data['atom_types']), 3))
            for frame in range(data.get_nframes()):
                if tutorial:
                    used_atoms_dict = {0: 'Si'}
                else:
                    used_atoms_dict = atoms_dict
                deepmd_system = ase.Atoms(symbols=[used_atoms_dict[num] for num in data['atom_types']], positions=data['coords'][frame],
                                          cell=data['cells'][frame], pbc=True,
                                          calculator=DP(model=f'{model_path}/{model_file}'))
                energy[frame], force[frame] = deepmd_system.get_total_energy(), deepmd_system.get_forces()
        else:
            energy, force, virial = model.eval(data['coords'], data['cells'], data['atom_types'])
            energy = energy[:, 0]
        E_pred.append(energy)
        F_pred.append(force)
        E_true.append(data['energies'])
        F_true.append(data['forces'])
    E_pred, F_pred = np.concatenate(E_pred), np.concatenate(F_pred)
    E_true, F_true = np.concatenate(E_true), np.concatenate(F_true)
    print('Energy RMSE (eV):      %.4e' % ((E_pred - E_true) ** 2).mean() ** 0.5)
    print('Energy RMSE/atom (eV): %.4e' % (((E_pred - E_true) ** 2).mean() ** 0.5 / data['coords'].shape[1]))
    plot_gen(F_true.flatten(), F_true.flatten(), labels=['F [eV/A]', 'F [eV/A]'], label='Real', color='k', show=False)
    plot_gen(F_pred.flatten(), F_true.flatten(), labels=['F [eV/A]', 'F [eV/A]'], scatter=True, label='Predicted',
             show=False, save=True, savefig=f'{model_path}/parity_plot_f_ase_{is_ase}.png')
    plt.clf()
    plot_gen(E_true.flatten(), E_true.flatten(), labels=['E [eV]', 'E [eV]'], label='Real', color='k', show=False)
    plot_gen(E_pred.flatten(), E_true.flatten(), labels=['E [eV]', 'E [eV]'], scatter=True, label='Predicted',
             show=False, save=True, savefig=f'{model_path}/parity_plot_e_ase_{is_ase}.png')

The training process is the usual use:
dp train input.json
dp compress -i graph.pb -o graph_compressed.pb

[wanghan-iapcm]

what was atoms_dict
how did you call plot_e_f_deviation

[wanghan-iapcm]

plz provide the code version

[biglinn]

Hellow!
I also want to use DP model with ASE package. Could you tell me how to use ase.Atoms with the DP calculator, or where is the interface program?

[anyangml]

from deepmd.calculator import DP
dp_calc = DP("XXX.pth")
atom.set_calculator(dp_calc)

[biglinn]

Thanks!