Can you try running with a smaller tilmestep to see if you
that makes your model more stable? I can’t think of
any reason why only switching PPPM vs Ewald would
make a SHAKE water model crash. Unless it is some
source of bad dynamics in your model (e.g. high temperature,
too-big timestep, etc), that will just cause occasional, random
problems). Also what is your pair style cutoff? You should only
get that error if an atom has moved further than the cutoff
away from another atom in the same water molecule.
Finally, if you use a high
accuracy setting and run for a few timesteps (thermo
every step), do you get (nearly) identical answers for NPT
dynamics with both Ewald and PPPM?
I switched from a 1 fs timestep to a 0.1 fs timestep and it made it a tiny bit further (49 ps rather than 30 ps); however, I still encountered the shake error.
For the record - this is SPC/E water at 298.15 K and 1 bar - so it isn’t a super extreme set of conditions or anything out of the ordinary.
My pair_style cutoff is 10.5 Angstroms (as is the cutoff between real_space and k_space as well). I’ve also tried slightly shorter cutoffs (7.6, 9.8); however, I can’t make it much longer as my box side length is about 21.6-21.9 angstroms (based on the volume fluctuations).
As for your high-accuracy suggestion, I moved up to a 1e-6 accuracy for the kspace style and did a comparison. I plotted the ewald and pppm volume fluctuations on top of each other each timestep for the first 10,000 timesteps (vol_flucts_timestep.png) and over the first 2,000 (vol_flucts_zoom.png). As you can see - at early times they agree very well (to decimal place precision); however they do eventually deviate from one another (as I think we would expect).
I was just wondering if it was something to do with how the shake molecules are handled across periodic boundary conditions - is it possible that something is not working correctly with that for ewald summation? That is the only thing that I could think of that would cause such a big jump if it really is a moving out of the cutoff range type of situation? I should mention - I don’t see crazy motions immediately before the crash when I visualize the trajectory.
I really appreciate your help troubleshooting this problem!
Both Ewald and PPPM know nothing about periodic BC.
I’ll ask Stan (CCd) to see if he can reproduce this or
has an idea what could go wrong. He know the long-range
solver details better than I.
Thanks for taking a look into this. I look forward to hearing from Stan.
Thanks so much,