using PBC for npz dataset

[ken1944]

Hello,

I tried finding this in the documentation and discussions but wanted to confirm. I'm using NequIP and training a model with a dataset that uses "pbc". My data format is npz and I'm specifying a pbc array in the npz file like pbc = np.array([19.5894, 19.5542, 19.5796], dtype='f') with the x,y and z dimensions. I wanted to confirm if this was the right way to mention the pbc since the output file (with and without pbc) shows pbc in the form pbc=[nframe,3]

Thanks for the help!

[Linux-cpp-lisp]

Hi @ken1944 ,

Thanks for your interest in our code!

We distinguish between pbc (whether each of the x,y,z directions is periodic) and cell, which gives the box that is periodic as a (3, 3) shape tensor whose rows are the three xyz cell vectors. So you'd want to set pbc = (True, True, True) and cell = [[19.57, 0, 0], [0, 19,57, 0], [0, 0, 19.57]], basically.

[ken1944]

Hi @Linux-cpp-lisp,

Thanks for your response! The model seems to train with pbc now. I also went ahead and added pbc and cell under npz_fixed_field_keys as mentioned in full.yaml

Thanks again!

[Linux-cpp-lisp]

👍

Yes, exactly: when all examples in our dataset have the same pbc/cell, that's what fixed_fields is for.