Langevin Integrator incompatible with RE-EDS Ensemble

[robinwuff]

Trying to simulate a replica-exchange enveloping-distribution sampling ensemble while using the LangevinIntegrator as a sampler causes various overflows after a few exchanges. The trajectories' positions become NaN after the overflow and the positions sampled before that equate to physical nonsense (see the following graph).

I believe this happens due to the BBK-style implementation using the positional difference between the current and the last sampled point. In the step following an exchange, this difference in position can get large and will shoot the sampler up the walls of the potential until after a few exchanges, the energy values simply get to large for calculation.

I don't know whether LangevinIntegrator and RE-EDS can be made to work together. If not, a warning would be appreciated for those who try and use that combination.

(I am aware now of the LangevinVelocityIntegrator, but that requires its separate issue)