Application of the restart command in shear simulation

I am using LAMMPS version 22 Dec 2022 on a CentOS system to conduct a shear box deformation simulation. During the simulation, I used the write_restart command to output an atomic configuration at a specific moment as restart_shear.equil, and then continued the simulation for an additional 100,000 steps. Subsequently, I read the output restart_shear.equil from a different file and continued the simulation for another 100,000 steps.

I monitored the epair changes obtained using thermo_style for both 100,000-step simulations (the original one and the one reading the restart file). Initially, the two were almost identical, but after approximately 50,000 steps, differences began to emerge.
Figure_1

Is this a normal phenomenon due to finite computational precision, or have I made a mistake in my settings somewhere?

Attached:
Original LAMMPS input file

units 		    lj
dimension 	    3
boundary 	    p p p
atom_style 	    atomic

lattice 	    fcc 1.2
region 	        box block 0 8 0 8 0 8
create_box 	    2 box
create_atoms 	1 box 
group           smaller_atoms id 1:19652:5
set 		    group smaller_atoms type 2

change_box 	    all triclinic

mass            1 1.0
mass            2 1.0

neighbor 	    1.0 multi
neigh_modify 	delay 0
comm_modify 	mode multi

pair_style 	    lj/cut 2.5
pair_coeff 	    1 1 1.0 1.0 2.5
pair_coeff 	    2 2 0.5 0.88 2.5
pair_coeff 	    1 2 1.5 0.8 2.5
pair_modify     shift yes

velocity 	    all create 1.0 882564 dist gaussian rot yes

timestep 	    0.005
fix             eq all nvt temp 1.0 1.0 $(100.0*dt)

thermo 	        1000
thermo_style 	custom step temp epair etotal press vol pxy

run 		    200000
unfix			eq

velocity        all ramp vx 0 $(0.002*ly) y 0 $(ly) sum yes units box
fix             1 all deform 1 xy erate 0.002 remap v
fix             2 all nvt/sllod temp 1.0 1.0 $(100.0*dt)

run				200000

write_dump		all custom dump_config_shear.colloid id type x y z
write_restart	restart_shear.equil

run				100000

Restart LAMMPS input file (restart lmp file)

read_restart    ./restart_shear.equil

neighbor 	    1.0 multi
neigh_modify 	delay 0
comm_modify 	mode multi

timestep	    0.005

fix             1 all deform 1 xy erate 0.002 remap v
fix             2 all nvt/sllod temp 1.0 1.0 $(100.0*dt)

thermo 	        1000
thermo_style 	custom step temp epair etotal press vol pxy

run		        100000

Yes. Anything that changes the order of the atoms in the local arrays will cause some tiny differences in the forces for some atoms and that difference will exponentially grow over time.
Restarting usually affects when the next neighbor list build happens and that will reorder some atoms and thus can cause differences when summing the forces. Similar things would happen if you change the number of MPI processes or use an executable compiled with a different compiler or with different optimization settings or run on a different platform.

The case you are presenting is looking pretty good. In many simulations this divergence of the energy becomes visible already after about 10,000 steps.