Compatibility of Langevin w/PBC

[Natha Hayre]

I'm almost positive that combining the Langevin integrator with any one of the periodic nonbonded methods results in a severe systematic drift in the coordinates of the solute molecule. I repeated several runs using OpenMM versus routines from a different MD package, and there are significant differences. Then, I switched to the Andersen thermostat with the Verlet Integrator, and everything seems okay.

Unfortunately, I really like the ability to modify the temperature of the Langevin thermostat, for the purposes of annealing. The Andersen thermostat has no such control.

Is the problem with Langevin dynamics and periodic boundary conditions a bug, or is this to be expected, or might I be using the interface incorrectly?

[Peter Eastman]

I've never seen any effect like what you're describing. Can you provide a test case?

Remember that a Langevin integrator applies random forces to the particles that are uncorrelated, and hence need not add up to 0. It will therefore cause each molecule's center of mass to move. This has nothing to do with periodic boundary conditions, and in fact is a fundamental feature of Langevin dynamics. Is that what you're talking about? One usually combines Langevin dynamics with a CMMotionRemover. That will prevent the system as a whole from drifting, but will still allow molecules to drift relative to each other.

Peter

[Natha Hayre]

The clue in my case was that the same system, with or without water molecules, would have the expected diffusive behavior with either of the NoCutoff and CutoffNonPeriodic methods, but shows a systematic, slow velocity in one particular direction with any the periodic non-bonded methods. I believe it is the same direction every time, with only displacement and not rotation. The velocity is on the order of 0.1 Angstrom per picosecond. I haven't yet looked into the temperature dependence of this effect, and I'm not sure if the waters are also drifting in the same direction. I know it's reproducible with any system I try with Langevin dynamics and periodic boundary conditions of any kind.

I can construct and send you a test case soon, but I just want to emphasize that this is definitely systematic, non-diffusive motion, and clearly not a natural outcome of Langevin dynamics.

[chris Neale]

Are you using a different number every time you seed the integrator? I experienced this in gromacs with many continuations due to cluster runtime limits of 24h and if I specified a seed (instead of asking for a new random seed every time it started) then there were very bizarre artifacts. These artifacts were worse in serial because there was some randomization of which atom got which "random" kick in a parallel simulation. I barely use openmm, but thought you might be running into the same thing here.