The MSST randomly break down with deep potential

LAMMPS
,
1

When I was using deepmd potential to siumlate the RDX MSST, I found it randomly break down in some shock velocity. And I found all the thermo parameters was appropriate, all the distance was proper. Just all of sudden, the direction of the shock was turned into nan.The 90 shock velocity crashed, but the 96 shock velocity was running properly. Is it caused by the intensive pressure fluctuation?But why is it not crashing in even the higher shock velocities?
This is the thermo properties

image
image2250×66 6.71 KB

image
image2276×134 11.4 KB

here is the input file


variable        FREQ            equal 25
variable        V_SHOCK         equal 96
variable        THERMO_FREQ     equal 10
variable        NSTEPS          equal 40000
variable        DUMP            equal 10

units           metal
boundary        p p p
atom_style      atomic

read_data       conf.lmp
mass            1 1.008000
mass            2 12.011000
mass            3 14.007000
mass            4 15.999000

pair_style      deepmd ../graph.000.pb ../graph.001.pb ../graph.002.pb ../graph.003.pb out_freq 25 out_file model_devi.out
pair_coeff      * *

neighbor        1.0 bin

min_style cg
minimize      1.0e-6 1.0e-6 10000 10000

#velocity       all create 300 234567
timestep 0.0004
# NPT fix
fix     npt all npt temp 300 300 0.01 iso 1.0 1.0 0.1
thermo_style custom step temp press vol ke pe etotal
thermo ${THERMO_FREQ}
run 10000
unfix npt
write_restart npt.restart

reset_timestep 0

# MSST fix
timestep 0.0002
fix     msst all msst  x ${V_SHOCK}

fix_modify msst energy no

variable dhug equal f_msst[1]
variable dray equal f_msst[2]
variable lgr_vel equal f_msst[3]
variable lgr_pos equal f_msst[4]

thermo_style custom step temp press vol ke pe etotal lx ly lz pxx pyy pzz  &
             v_dhug v_dray v_lgr_vel v_lgr_pos f_msst
thermo ${THERMO_FREQ}
dump            dpgen_dump all custom 25    all.lammpstrj id type x y z

run     ${NSTEPS}
2

Does your intended phenomenon really occur at a constant temperature of 300K and a constant pressure of 1 bar? If not, then it is probably not accurate to simulate it with an NPT integrator.

I happen to know that machine learning MD doesn’t do very well with pressures for many molecular systems because the electrostatic long-range interactions are missing. But I can’t know whether that applies to your particular simulation.

Also note that, in general, an MD simulation that is generally unsuitable (in terms of unrealistic potentials) will often nevertheless run for reasonably long durations and only occasionally crash. (After all, if it was unsuitable enough to crash under any circumstance, it would have crashed very quickly after starting for just about any configuration and the problems would be quickly detected and sorted out.)

3

Thank you for your detailed reply. The NPT running was nice, it was the MSST running after the NPT equilibrium that break down. I think the potential was not that bad for that the msst running was not crashing using my own python MSST code.