Reproducing MD run

[burakgurlek]

Hi,

I am trying to test certain settings of MLFF by using the same MD trajectories. I did use the same random seed for both calculations, but after 10th step the trajectories differ. I use Nose-Hoover Chain thermostat.

Does the same trajectory possible with the selected thermostat? I guess different MLFF settings should not modify the trajectory.

Best Regards,
Burak

[ferenc_karsai]

For clarification, the tiniest change in a single step will create a snowball reaction, so it is more or less impossible to reproduce runs for thousands of MD steps. Hence if one wants to check if the trajectory is correct it is always better to look for example at the pair correlation function and see if it is the same as the one for the reference trajectory.

But if you still want to reproduce the trajectory, here are some remarks:
If you use anything where ab-initio calculations are involved (ML_MODE=TRAIN) then reproducibility is only possible until a few steps even if you use a random seed. This is even the case if you use a single processor, since the FFT maps are different at every run. On multiple processors this is much worse of course.

The best chance of being able to reproduce a few tens or hundreds of steps is using ML_MODE=RUN and a single processors.