Dear Lammps users,
I’m trying to repeat some simulations from the literature work using respa with eam potentials. But it seems that the
was this simulation performed with LAMMPS? what are the specific
directions how r-RESPA was applied, what cutoff ranges and what
force decomposition algorithm (with inner, middle, and outer keyword in respa) has not been implemented for eam
potential in lammps, but implemented for LJ potential. I wonder why? Is it because technically it is very difficult or even impossible to decompose forces from eam potential due to its manybody feature, where the atoms in one layer will also affect the force contribution from atoms in another layer?
it is technically difficult, and the benefit is limited.
1) for 3-tier inner/middle/outer pairwise interactions to work
efficiently, you need a significant separation and geometry changes
between the distance ranges, and that requires using a rather long
outer cutoff. EAM potentials, have a quite short cutoff, so that won't
make sense. you need some overlap region to smoothly switch between
the distance ranges so you retain sufficient energy conservation.
2) for 2-tier inner/outer decomposition, it is more likely to have a
(small) performance benefit from r-RESPA. in fact, i find this often
to be more efficient than the three level setup even for lj/cut, but
again, somebody needs to program the switching function.
3) the eam force computation is a 2-step process. you have to walk the
neighbor list twice. you first compute the embedding energy
contribution from the neighboring atoms, make a reverse communication
(i.e. reduction) and then compute the embedding term contribution and
do a forward communication before computing forces and energies from a
pairwise term and the embedding term. so you need that first step in
any case, or figure out a way to decompose and store those
contributions for each r-RESPA level in addition to the forces and
avoid having to recompute them. LAMMPS is not set up for that
in short, it would be a lot of work, but it is not obvious, if there
would be a sufficient benefit for typical EAM potentials the way
LAMMPS is currently programmed. in today's HPC environment, it is much
easier to obtain access to additional computational resources, and
computers are much more capable, than they were when r-RESPA was a big
nowadays, you can probably gain almost as much performance, if not
more by simply using MPI+OpenMP or MPI+GPU MPI+KOKKOS or
MPI+USER-INTEL with plain verlet time integration.