Project

General

Profile

Feature #2944

Roadmap for thermostats / barostats in new propagation/integration scheme

Added by Michael Shirts 4 months ago. Updated 4 months ago.

Status:
New
Priority:
Normal
Category:
mdrun
Target version:
Difficulty:
hard
Close

Description

We are in the process of putting thermostats and barostats into the proposed new modular integrator framework.

Thermostats we are planning to put in, roughly in order:
  • Langevin integrators. We are planning on implementing the framework presented be Leimkuhler et al. (https://aip.scitation.org/doi/10.1063/1.4802990, also discussed in https://pubs.acs.org/doi/abs/10.1021/jp411770f). Note that this would not just be a SINGLE Langevin integrator, but the framework to swap around orders of the different steps (velocity/position/Ornstein-Uhlenbeck process), so that a number of different algorithms could be implemented easily.
  • vrescale of Bussi/Parrinello: the simplest scaling thermostat that rigorously gives NVT behavior in an ergodic system.
  • weak coupling / Berendsen: we may include this solely for equilibration, but perhaps we should eliminate this entirely in favor of v-rescale. If remaining, maybe bump up warnings even more about using only for prepping systems.
  • Andersen: this would be relatively straightforward to put in, though may not be needed if Langevin is supported. Worth discussing
Thermostats we would very much prefer NOT to put in:
  • Nose Hoover / Nose-Hoover with chains. There have been a number of papers that have shown that in systems with decoupled degrees of freedom, Nose-Hoover is nonergodic, and has other weird artifacts. The errors appear to be negligible for most fluids - but NOT negligible when there are decoupled systems of equations with some subsystems containing relatively few variables (such as gas phase molecules, alchemical transformation endstates). See https://journals.aps.org/pre/abstract/10.1103/PhysRevE.52.2338, and https://aip.scitation.org/doi/10.1063/1.2989802. I would argue this is the time to get rid of this formalism. I am open to discussing, though!
Barostats to put in:
  • We are planning on supporting a Monte Carlo barostat.
  • We can probably support Parrinello-Rahman with Langevin thermostatting of the barostat DOF (Nose-Hoover thermostatting should not exist, see above). Though It's not entirely clear that it's needed. The dynamics are as fictitious as MC, really. It's worth a discussion.
  • Berendsen barostat SOLELY as equilibration/system prepping. Must be heavily warned.

History

#1 Updated by Mark Abraham 4 months ago

We mooted at the dev telco that we should invite dev+user feedback in a post that describes the issues known with N-H (Michael to write some words to help with this), also invite comment from knowledgeable people who might not see gmx-users (e.g. Martyna, Tucker, etc.), and forecasts deprecating it in GROMACS 2020. We will not implement N-H in the new framework for GROMACS 2020. If we learn something new, we might consider implementing N-H in the new framework some time after we release GROMACS 2020.

GROMACS 2020 should warn for berendsen thermostat and barostat (if it doesn't already)

Also available in: Atom PDF