integrated heat flux diverging

Is there any way to figure out why, in 16 identical except for random seed runs, the heat flux diverges linearly in a few cases?

I'm creating the initial velocity distribution with "gaussian" and "mom yes", and then running some fix nvt (equilibration) followed by fix nve (flux data collection), which I thought would be enough to prevent overall translation. any given heat flux autocorrelation plot for a 5e4 atom system (and its integral) is rather noisy, but averaging over 16 random seed instances smooths it out to a nice plateau. However, I'm disturbed that sometimes even though the decorrelation time appears to be ~30 ps for my system, in some runs the integrated value just increases linearly for 600 ps. Does anyone have any suggestions for what might be going wrong or possible fixes, or does this just sounds like normal behavior?

                Noam

This is pure speculation, but the velocity command uses a rather low quality pseudo random number generator (RanPark). It is needed for the loop geom option.
However, when removing that option, one can also use the (better) other pseudo random number generator in LAMMPS.
Please try with the attached files, that do the replacement and provide a new command “velocity/mars” that uses the RanMars random number generator and let us know, if that has an impact.

Thanks,
Axel.

velocity_mars.tar.gz (7.17 KB)

Hi Axel - I tried your velocity/mars patch, but it unfortunately made no difference. I’m currently suspecting some inherently difficult dynamics (comparing my autocorrelation to the analysis method in https://github.com/lorisercole/thermocepstrum).

Noam