Many-body heat flux correction

[rfhari]

Hi @yury-lysogorskiy

I am currently exploring to use the PACE framework to predict the thermal conductivity of AlGaN alloys using the Green-Kubo method (by performing equilibrium molecular dynamics simulations). So I was just wondering how are multi-body heat flux corrections (the atomic virial terms) handled in the implementation of PACE interface with LAMMPS? should I write a wrapper based on the the forces predicted by PACE?

few relevant papers (maybe):

• Heat transport in liquid water from first-principles and deep neural network simulations
• Force and heat current formulas for many-body potentials in molecular dynamics simulations with applications to thermal conductivity calculations

[yury-lysogorskiy]

Hi @rfhari , thank you for using PACE!

We explicitly provide "pair" forces for each atom to LAMMPS (https://github.com/lammps/lammps/blob/develop/src/ML-PACE/pair_pace.cpp#L223) so, I would assume that atomic virials also should be computed by LAMMPS correctly.

Please, check it and let us know if it work.

Regards,
Yury

[brucefan1983]

I think ACE does not have the required 9-component per-atom virial for correctly calculating the heat flux yet.

In terms of code, one must use cvflag_atom in LAMMPS interface.

For theoretical background, one can check this review article that demonstrates how MTP (very similar to ACE) is failing to produce the correct heat flux: Molecular dynamics simulations of heat transport using machine-learned potentials: A mini-review and tutorial on GPUMD with neuroevolution potential. Therefore, it is not sufficient to provide pair forces for each atom to produce the correct atomic virials for heat flux calculation. More efforts are needed to make heat flux right.