A PhD student here noticed that the results of some granular simulations he was running seemed to vary depending on the skin distance selected. After looking into it a bit more, the frequency at which the neighbour lists are rebuilt affects the results when this should not be the case. I attach two files, a simple input script (in.test) and a data file of particle positions and velocities, which demonstrate the problem. This script compares two cases: rebuilding the neighbour lists based on the usual skin distance check, and rebuilding the neighbour lists on every timestep. There are substantial differences between the output dumpfiles after 100000 timesteps. I’ve been unable to track down the reason for this disparity, but I have established that the problem occurs using a serial build of the latest version of LAMMPS and is unaffected by shear history.
I’ve also noticed that a similar divergence of results occurs when running the 3D Lennard-Jones melt example given in examples/melt. If the following two cases are compared:
neigh_modify every 1 delay 0 check no
neigh_modify every 1 delay 0 check yes
a divergence occurs in the thermo output within 1000 timesteps. Since the same machine and random number seed is used, my understanding is that these results ought to be identical. Perhaps these two issues are unrelated, but the former is causing a problem as some outputs of our granular simulations appear to be highly sensitive to the skin distance.
Dept. of Civil and Environmental Engineering
Imperial College London, SW7 2AZ
in.test (995 Bytes)
Particle_Coordinates.lj (77.2 KB)