DeepWannier NN predicts 0 global dipole only, no atomic dipole

[alikhamze]

Hello,

I trained a DeepWannier neural network on a small set of training data (structures generated using MD at 500 K, then Wannier center calculations done using quantum espresso) as a proof-of-concept before generating a larger training set for an actually accurate model. The training completed successfully, and I then froze and compressed the model.
To test the model, I generated additional structures for the same system (again using MD, at 300 K this time) to apply the model to.

I performed the inference following the script from the Princeton workshop over the summer: https://github.com/CSIprinceton/workshop-july-2022/blob/main/hands-on-sessions/day-2/4-deep-wannier/dipole_distribution.py

However, when I infer from my model, I either get 0 or no output at all. When I ask for a global dipole moment only, I get 0:

When I ask for the atomic dipoles as well, the output is blank:

When I use the model distributed in the Princeton workshop with the structures provided, I do get non-zero dipole moments predicted by the model, so I don't think my DeepMD installation is the source of the problem. However, in case the information will be helpful, I am using DeepMD v2.1.5.

Here is my training input file (converted from json to txt to be uploaded to github).
input.txt

Does anyone have any idea why my toy model won't predict a non-zero dipole moment?

Thanks,
-Ali

[salinelake]

Hi Ali,

Can you provide a minimal example of the training set?

[alikhamze]

Hi Pinchen,

Attached is the data from the first 10 frames. Let me know if anything looks wrong?

Thanks for taking a look!

sample_data.tar.gz

[salinelake]

I trained a model with the data you provide. The model looks alright. You can test your model directly against the training data by

from deepmd.infer import DeepDipole 
import numpy as np
import dpdata
import os

trainset = [
        "./data"
        # "/home/ali/calculations/bn/train_deepwannier/data/processed",
]
 
## load all training data
model_path = './dipole.pb'
model = DeepDipole(model_path)
## evaluate the trainset
dipole_ref = []
global_ref = []
dipole_pred = []
global_pred = []
for data_dir in trainset:
    print('loading ', data_dir) 
    cells = np.load(os.path.join(data_dir,'set.000/box.npy'))
    coords = np.load(os.path.join(data_dir,'set.000/coord.npy'))
    atypes = np.loadtxt(os.path.join(data_dir,'type.raw')).astype(int)
    dipole_ref.append(np.load(os.path.join(data_dir,'set.000/atomic_dipole.npy')).reshape(-1,3))
    global_ref.append(np.load(os.path.join(data_dir,'set.000/dipole.npy')).reshape(-1,3))
    _dipole_pred = model.eval(coords, cells, atypes)
    dipole_pred.append(_dipole_pred.reshape(-1,3))
    global_pred.append(_dipole_pred.sum(-2).reshape(-1,3))

dipole_ref = np.concatenate(dipole_ref)
dipole_pred = np.concatenate(dipole_pred)
global_ref = np.concatenate(global_ref)
global_pred = np.concatenate(global_pred)
##
print(global_ref.shape)
print(global_pred.shape)

I guess the fact you get either get 0 or no output following the script from the Princeton workshop is due an incorrect atypes:

atypes = np.array(atoms.get_atomic_numbers()) - 1
dipole = dipole_model.eval(atoms.get_positions(),atoms.get_cell(),atom_types=atypes)

Please check "positions" and "cell" feeded to dipole model has the correct shape.
And make sure atom_types is given consistently with your training data.

[alikhamze]

Finally got a chance to try this and double check things, this worked! Thank you, Pinchen!