ScaleShiftMACE: torch.autograd.grad() outputs parameter using inter_e instead of total_energy?

[JSLJ23]

I have a question behind the logic / reasoning of how the forces are computed in the ScaleShiftMACE variant of MACE.
I notice in the get_outputs() function, the energy term passed to this function is the inter_e which is then used by the compute_forces() function to call torch.autograd.grad() for the gradients with respect to the positions.
I have 2 questions regarding this:

1. What does this mean physically when the inter_e is used instead of the total_energy?
2. Is there a physical meaning to the inter_e and why back-propagating with it in the context of ScaleShiftMACE would lead to forces?

[pobo95]

To fit a potential to the binding energy.

inter_e is E0 (energy of isolated atoms).

[ilyes319]

The inter_e corresponds to energy of the system without the isolated energy. This quantity is very physical, as isolated energies are just arbitrary shifts that does not depend on the geometry. The reason why we backpropagate through it, is to avoid floating point precision issues, as the total energy is usually a large number for most DFT codes.

[JSLJ23]

Thank you for your answer. Could you expand a bit more on the "isolated energies" you were referring to previously? Does it refer to the single point energy of an isolated atom?

[ilyes319]

Yes exactly.