Kokkos user-reaxc variable problem

All,

I’ve compiled kokkos with user-reaxc (LAMMPS (23 Aug 2016)) but am getting an inconsistency between the reported PE from thermo and from:
compute ppa all pe/atom
compute pta all reduce ave c_ppa

Below is the portion log file showing the problem; notice how the 3rd column increases while the 2nd does not:

Step PotEng c_pta Temp v_staxx v_stayy v_stazz
0 -174.50649 -174.50649 0 -0.0014747196 -0.0014747196 -0.0014747196
1 -174.50649 -349.01298 1.3766098e-28 -0.0029494392 -0.0029494392 -0.0029494392
2 -174.50649 -523.51947 5.5072519e-28 -0.0044241588 -0.0044241588 -0.0044241588
3 -174.50649 -698.02596 1.2389307e-27 -0.0058988784 -0.0058988784 -0.0058988784
4 -174.50649 -872.53245 2.2023293e-27 -0.007373598 -0.007373598 -0.007373598
5 -174.50649 -1047.0389 3.4413241e-27 -0.0088483176 -0.0088483176 -0.0088483176
6 -174.50649 -1221.5454 4.9556372e-27 -0.010323037 -0.010323037 -0.010323037
7 -174.50649 -1396.0519 6.7454249e-27 -0.011797757 -0.011797757 -0.011797757
8 -174.50649 -1570.5584 8.8107492e-27 -0.013272476 -0.013272476 -0.013272476
9 -174.50649 -1745.0649 1.1152048e-26 -0.014747196 -0.014747196 -0.014747196
10 -174.50649 -1919.5714 1.3767987e-26 -0.016221916 -0.016221916 -0.016221916

I’ve attached a minimal input script to reproduce the problem and the complete log file.

My LAMMPS command is:
mpiexec -np 1 omplace -nt 1 -vv lmp_kokkos_cuda20160826 -in in.script -v g gpu -echo both -k on t 1 g 1 -sf kk -pk kokkos newton on neigh half binsize 12 comm no

Thanks.

-Ben

ffield.reax.C2013 (1.97 KB)

in.script (1.77 KB)

log.gpu.lammps (4.59 KB)

Hello LAMMPS users, I am posting this to solve one dilemma: which way is better to simulate chemical reactions. One way is ReaxFF of course.

Another way is using non-reactive force field and fix bond/create which has only distance criteria to form chemical bond , therefor user needs to monitor energies of particles prior to reaction (based on DFT energies of activation and dumping energies of reacting particles to decide to keep or break back the bond which was just created by fix bond/create). Neverless I see here problem what exactly is energy of activation in LAMMPS simulation (pe, ke, pair/interaction etc).

Does anyone have experience with both of these methods or has anyone hear about this second method ? Moreover what is your opinion about this second method from scientific point of view , how good is to be accepted like “valid method” ?

Moreover what about publishing in papers, is second method good enough to be accepted for paper ?

Has anyone seen this second method in some published paper ?

How slower is ReaxFF compared to non-reactive force fields ? (if someone can tell from experience).

Thank you all and greetings .

  Hello LAMMPS users, I am posting this to solve one dilemma: which way is
better to simulate chemical reactions. One way is ReaxFF of course.

this is an ill-posed question. there rarely is an absolute "better".
at best you'll have "better for purpose X".

   Another way is using non-reactive force field and fix bond/create which
has only distance criteria to form chemical bond , therefor user needs to
monitor energies of particles prior to reaction (based on DFT energies of
activation and dumping energies of reacting particles to decide to keep or
break back the bond which was just created by fix bond/create). Neverless I
see here problem what exactly is energy of activation in LAMMPS simulation
(pe, ke, pair/interaction etc).

bond create is mean to be used for a specific class of use cases. what
you describe goes beyond it, and would require things like a monte
carlo acceptance criteria similar to what fix bond/swap or fix gcmc do
(for different purposes).

   Does anyone have experience with both of these methods or has anyone hear
about this second method ? Moreover what is your opinion about this second
method from scientific point of view , how good is to be accepted like
"valid method" ?

this is a question that is best answered with a thorough research of
the published literature.

   Moreover what about publishing in papers, is second method good enough to
be accepted for paper ?

again, a literature search will give a better answer than just the
opinion of a bunch or (random) people.

    Has anyone seen this second method in some published paper ?

i repeat, do a survey of the literature.

   How slower is ReaxFF compared to non-reactive force fields ? (if someone
can tell from experience).

did i already tell you to do improve your searching skills (rather
than asking others to do your job for you)?
you should have found this (on the LAMMPS homepage):

http://lammps.sandia.gov/bench.html#potentials

axel.