Apparently, it is more likely to happen if the simulation has been running for little time (50-100fs). Also commented in https://github.com/pandegroup/openmm/issues/1483