Multi-body heat flux calculation using ACE+LAMMPS

[rfhari]

Hi,

I have a question regarding calculating multi-body heat flux using ACE. I think this question does not really belong here but I cannot think of another place to ask.

I am using Atomic Cluster Expansion (ACE) potential + LAMMPS to study heat transport in solid materials. Currently I'm applying it to AlGaN alloys in equilibrium/Green-Kubo methods. I'm interested in adopting the heat flux corrections noted by Surblys et al to address the atomic stress virial in many body potentials, specifically the ACE potential. But I'm not fully sure how to implement it for heat flux calculated by ACE as I haven't come across any relevant literature/work implementing the same (even for other MLFF like SNAP, MTP etc.)

So I was wondering if anyone here have tried using ACE to study thermal transport problems using Green-Kubo?

[gabor1]

I transferred this to the ACE repo. which ACE package are you using, is it PACE ? I am not familiar with the details of the implementation, but the stress tensor that is sent back to LAMMPS is a very likely a sum of atomic terms (because everything in ACE is a sum of atomic terms), so in principle all you need is that atomic stress tensor before is summed up. Others here may have done G-K with ACEs.

[bernstei]

I am saying that when you have access to a LOCAL site energy and a site energy gradient, then all other terms become canonical and straightforward to define. This has nothing to do with ML potentials.

I agree with that. But the wealth of publications about how to compute the heat flux (or even local stress) suggests that maybe not everyone does. Although, maybe the fundamental issue is really whether different definitions of the local site energies (and corresponding heat fluxes) lead to different thermal conductivities.

For non-local site energies the issue is more subtle but not insurmountable.

I don't see how this can possibly be true, or at least given what I think you mean by non-local :) My understanding is that, by definition, the heat flux is about the transport of a locally defined energy. If the energies are truly non-local, I don't see how the heat flux concept is even well defined.

[cortner]

Ok - fair point about nonlicality, one needs some emergent locality here, so let's leave that aside.

I will take a bet that in the limit of infinite computational domains and infinite time the heat flux is independent of the definition of local energy. But different definitions will give different speeds of convergence- or said differently better or worse accuracy for fixed computational budget.

I should substantiate this with a reference but don't have the time and energy to look for it. If people here think this is genuinely controversial then I will find a rigorous reference. I believe this must be in decade-old statmech results.

If not then it's time to write it up.

[bernstei]

I will take a bet that in the limit of infinite computational domains and infinite time the heat flux is independent of the definition of local energy.

I think this is surely true, but I'm not sure how widely accepted it is. It doesn't help that because the heat flux is constructed from a local (site) stress, some of the discussion I'm remembering might refer to stress definitions rather than to the flux and corresponding GK thermal conductivity. Let me see what notes I have.

[wcwitt]

It seems clear we all agree the core ingredients are available from LAMMPS, but also that we may need some experimentation to probe some of claims in the literature for ourselves. Perhaps the OP will reply (maybe a tag will help @rfhari) but in the meantime do any of you have need for these calculations in the short term?

Alternatively, I know @nickhine (hopefully this tag works) would like to compute G-K thermal conductivities with MACE, so perhaps one of his examples could become a test case.

[rfhari]

Thank you all for the interesting discussions here. I am just starting to explore the exact details of papers on heat-flux corrections (accounting for multi-body, bond/dihedral etc.), as @bernstei mentioned, and I think the details discussed here will be helpful. I will checkout the LAMMPS and PACE links shared by @wcwitt and get back if needed. Also, yes, it would be really great if @nickhine can share his thoughts on computing G-K thermal conductivities with MACE.