Shift in energy and forces with lammps ?

[mcourtiel]

Hi everyone,

When I perform molecular dynamic with a Neural Network Potential trained with DeepMD-Kit (v2.2.10) on lammps, I have found that when I extract the new structures/energy and forces to make a new set, and then dp test with the same NNP on this data set, I have a big shift in energy and the forces are totally different.

But I should find exactly the same energy ad forces as the dynamic was propagate thanks to this NNP...

There should be no units problems as I use "units metal" in lammps.
There is how I ask for the potential energy (pe) :
thermo_style custom step time etotal ke pe temp press vol density
thermo_modify flush yes
thermo 100
And how I ask for the coordinates and forces :
dump forces_dump all custom 100 traj.out id type x y z fx fy fz

Does someone understand what is happening ?

[njzjz]

And how I ask for the coordinates and forces :
dump forces_dump all custom 100 traj.out id type x y z fx fy fz

Could you also let me know if you also provided the box?

[mcourtiel]

That's the lammps input I use (the box is define in the RuS-dcbpy.lmp file :

units metal
boundary p p p
atom_style atomic

neighbor 1.0 bin
neigh_modify every 10 delay 0 check no

read_data RuS-dcbpy.lmp

pair_style deepmd graph-compress_1Ru-S.pb
pair_coeff * *

timestep 0.0005

dump dcd all dcd 100 RuS-dcbpy_400K.dcd

thermo_style custom step time etotal ke pe temp press vol density
thermo_modify flush yes
thermo 100

dump forces_dump all custom 100 traj.out id type x y z fx fy fz

velocity all create 200.0 12431001 dist gaussian

fix 1 all nve
fix 2 all langevin 400.0 400.0 $(100.0*dt) 18531001
fix 3 all momentum 1 linear 1 1 1

run 50000

[njzjz]

Did you provide the box when using dp test?

[mcourtiel]

Yes the box was provided. I have one data set with box.npy, coord.npy, energy.npy and force.npy.

The box.raw that I convert into .npy looks like :
16.5 0.0 0.0 0.0 16.5 0.0 0.0 0.0 16.5 (nframes time this line)

[njzjz]

It seems that fix langevin changes the energy and forces
https://github.com/lammps/lammps/blob/a0fcbc9b7112742cf4c89f8f73610b4f6ba24b74/src/fix_langevin.cpp#L994