Counter-intuitive Batch SCF Charge Convention

[MoleOrbitalHybridAnalyst]

My molecules all have charge=1 and I found I have to set charges to 1 / batch_size to get the correct results in batch mode. A simple example is as follows:

import torch

import dxtb
from dxtb.typing import DD

dd: DD = {"device": torch.device("cpu"), "dtype": torch.double}


numbers = torch.tensor(
    [
        [8, 1, 1, 1, 8, 1, 1, 8, 1, 1, 8, 1, 1, 8, 1, 1, 8, 1, 1],
        [8, 1, 1, 1, 8, 1, 1, 8, 1, 1, 8, 1, 1, 8, 1, 1, 8, 1, 1]
    ],
    device=dd["device"],
)
positions = torch.tensor(
    [
        [
                   [ -7.68281384,   1.3350934 ,   0.74846383],
                   [ -5.7428588 ,   1.31513411,   0.36896714],
                   [ -8.23756184,  -0.19765779,   1.67193897],
                   [ -8.13313558,   2.93710683,   1.6453921 ],
                   [ -2.95915993,   1.40005084,   0.24966306],
                   [ -2.1362031 ,   1.4795743 ,  -1.38758999],
                   [ -2.40235213,   2.84218589,   1.24419946],
                   [ -8.2640369 ,   5.79677268,   2.54733192],
                   [ -8.68767571,   7.18194193,   1.3350556 ],
                   [ -9.27787497,   6.09327071,   4.03498102],
                   [ -9.34575393,  -2.54164384,   3.28062124],
                   [ -8.59029812,  -3.46388688,   4.6567765 ],
                   [-10.71898011,  -3.58163572,   2.65211723],
                   [ -9.5591796 ,   9.66793334,  -0.53212042],
                   [ -8.70438089,  11.29169941,  -0.5990394 ],
                   [-11.12723654,   9.8483266 ,  -1.43755624],
                   [ -2.69970054,   5.55135395,   2.96084179],
                   [ -1.59244386,   6.50972855,   4.06699298],
                   [ -4.38439138,   6.18065165,   3.1939773 ]
        ],
        [
                  [ -7.67436676,   1.33433562,   0.74512468],
                  [ -5.75285545,   1.30220838,   0.37189432],
                  [ -8.23155251,  -0.20308887,   1.67397231],
                  [ -8.15184386,   2.94589406,   1.6474141 ],
                  [ -2.96380866,   1.39739578,   0.24572676],
                  [ -2.14413995,   1.48993378,  -1.37321106],
                  [ -2.39808135,   2.86614761,   1.25247646],
                  [ -8.26855335,   5.79452391,   2.54948621],
                  [ -8.69277797,   7.18061912,   1.33247046],
                  [ -9.28819287,   6.08797948,   4.03809906],
                  [ -9.3377226 ,  -2.54245643,   3.27861813],
                  [ -8.59693106,  -3.48501402,   4.65503795],
                  [-10.72627446,  -3.59514726,   2.66139579],
                  [ -9.55955755,   9.6716561 ,  -0.53106973],
                  [ -8.7077635 ,  11.28708848,  -0.59527696],
                  [-11.12540351,   9.87000175,  -1.44181568],
                  [ -2.70194931,   5.55490663,   2.9641866 ],
                  [ -1.60305656,   6.49854138,   4.07984311],
                  [ -4.39083534,   6.17898869,   3.18702311]
        ],
    ],
    **dd
)
charge = torch.tensor([0.5, 0.5], **dd)


# no conformers -> batched mode 1
opts = {"verbosity": 6, "batch_mode": 1}

calc = dxtb.Calculator(numbers, dxtb.GFN1_XTB, opts=opts, **dd)
result = calc.energy(positions, chrg=charge)

print(result.detach().numpy())

This gives the correct energy:

[-34.77268237 -34.77261976]

If I have charge = torch.tensor([1, 1], **dd) instead, I get the wrong energies:

[-34.02695165 -34.02703838]

which are the energies of the molecules as if their charges = 2.

So it seems that the charge tensor is multiplied by batch_size somewhere internally. I think this behavior is quite counter-intuitive and probably not the actually wanted behavior.

[marvinfriede]

Thanks for raising this issue. I confirmed that there is indeed something weird going on. I will take a look.

[marvinfriede]

I fixed it and created a bugfix release v0.1.1. You can download it from pip now. It will also be available via conda in a few days.