Using DFT-D3 dispersion#

TorchANI can use a built-in implementation of D3 dispersion for dispersion interactions. This is meant for using with functionals that are parametrized with dispersion in the first place.

Dispersion Energy: tensor([-0.003], device='cuda:0', grad_fn=<IndexAddBackward0>)
Force: tensor([[    -0.000,      0.000,      0.000],
        [     0.000,     -0.000,      0.000],
        [     0.000,      0.000,     -0.000],
        [    -0.000,     -0.000,     -0.000],
        [    -0.000,      0.000,      0.000]], device='cuda:0')
Dispersion Energy: tensor([-0.003, -0.003, -0.003, -0.003], device='cuda:0',
       grad_fn=<IndexAddBackward0>)
Force: tensor([[[    -0.000,      0.000,      0.000],
         [     0.000,     -0.000,      0.000],
         [     0.000,      0.000,     -0.000],
         [    -0.000,     -0.000,     -0.000],
         [    -0.000,      0.000,      0.000]],

        [[    -0.000,      0.000,      0.000],
         [     0.000,     -0.000,      0.000],
         [     0.000,      0.000,     -0.000],
         [    -0.000,     -0.000,     -0.000],
         [    -0.000,      0.000,      0.000]],

        [[    -0.000,      0.000,      0.000],
         [     0.000,     -0.000,      0.000],
         [     0.000,      0.000,     -0.000],
         [    -0.000,     -0.000,     -0.000],
         [    -0.000,      0.000,      0.000]],

        [[    -0.000,      0.000,      0.000],
         [     0.000,     -0.000,      0.000],
         [     0.000,      0.000,     -0.000],
         [    -0.000,     -0.000,     -0.000],
         [    -0.000,      0.000,      0.000]]], device='cuda:0')
Dispersion Energy: tensor([-0.013], device='cuda:0', grad_fn=<IndexAddBackward0>)
Force: tensor([[     0.000,      0.000,     -0.000],
        [     0.000,     -0.000,      0.000],
        [     0.000,      0.000,     -0.000],
        [    -0.000,     -0.000,     -0.000],
        [    -0.000,      0.000,      0.000]], device='cuda:0')

# To begin with, let's import the modules we will use:
import math
import torch

from torchani.potentials import TwoBodyDispersionD3
from torchani.grad import energies_and_forces

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
disp = TwoBodyDispersionD3.from_functional(
    functional="b973c", cutoff=math.inf, symbols=("H", "C", "N", "O")
).to(device)

coordinates = torch.tensor(
    [
        [
            [0.03192167, 0.00638559, 0.01301679],
            [-0.83140486, 0.39370209, -0.26395324],
            [-0.66518241, -0.84461308, 0.20759389],
            [0.45554739, 0.54289633, 0.81170881],
            [0.66091919, -0.16799635, -0.91037834],
        ]
    ],
    device=device,
)
znumbers = torch.tensor([[6, 1, 1, 1, 1]], device=device)
disp_energy, force = energies_and_forces(disp, znumbers, coordinates)
print("Dispersion Energy:", disp_energy)
print("Force:", force.squeeze())

# Dispersion can also be calculated for batches of coordinates
# (here we just repeat the methanes as an example, but different molecules
# can be passed by using dummy "-1" atoms in the species)
r = 4
coordinates = coordinates.repeat(r, 1, 1)
znumbers = znumbers.repeat(r, 1)
disp_energy, force = energies_and_forces(disp, znumbers, coordinates)
print("Dispersion Energy:", disp_energy)
print("Force:", force.squeeze())

# Different supported species can be passed down to the constructor
disp = TwoBodyDispersionD3.from_functional(
    functional="b973c", cutoff=math.inf, symbols=("H", "C", "N", "O", "S", "Fe")
).to(device)
# Here we change the species a bit to make a nonesense molecule
coordinates = torch.tensor(
    [
        [
            [0.03192167, 0.00638559, 0.01301679],
            [-0.83140486, 0.39370209, -0.26395324],
            [-0.66518241, -0.84461308, 0.20759389],
            [0.45554739, 0.54289633, 0.81170881],
            [0.66091919, -0.16799635, -0.91037834],
        ]
    ],
    device=device,
)
znumbers = torch.tensor([[6, 16, 1, 8, 26]], device=device)
disp_energy, force = energies_and_forces(disp, znumbers, coordinates)
print("Dispersion Energy:", disp_energy)
print("Force:", force.squeeze())