Hello,

Are you sure that your simulations' differences are only because of the thermostat?Have you ruled out other factors, such as seed numbers, which are used to initialize velocities? Note that if not explicitly set in the mdp, seeds are automatically assigned as a function of the computer's clock (date and time to the second) and therefore can be different for two otherwise identical ensembles. That's only one example of differences that can occur for the same simulation system. Keep in mind that Molecular Dynamics simulations are stochastic processes and therefore, the same simulation system with the same configuration is expected to give slightly different results in two different runs. This is totally normal, in fact, it is expected, and this is why we run the same simulation at least twice for validation. In this respect MD simulations are like wet lab experiments (you wouldn't trust an experiment performed only once with no replicates, would you?).

Having said the above, it is absolutely normal that two different algorithms give slightly different results, since V-rescale and Nose-Hoover approach the constant temperature problem from "opposite" sides.

V-rescale is, essentially, a modified version of the Berendsen weak coupling scheme, in the sense that it controls kinetic energy (and therefore temperature) by scaling velocity fluctuations.

The difference with "vanilla" Berendsen is that V-rescale applies an extra Wiener function and a random seed to address Berendsen's problems with velocity distribution.

On the other hand, Nose-Hoover has a more immediate effect, as it modifies the Hamiltonian itself rather than simply suppressing fluctuations. This leads to a proper Maxwell-Boltzmann distribution for velocity and therefore proper canonical ensemble.

When properly applied and equillibrated, both thermostats are expected to give "true NVT" (and NPT if pressure is also controlled). Therefore, you are expected to get almost (emphasis on 'almost') the same result for your simulation with only minor changes which, unless you're studying phenomena that are extremely dependent on kinetic energy and absolute temperature control.

Generally speaking, you should choose your algorithm based on the needs of the problem you want to investigate. Berendsen and v-rescale thermostats are good for equillibration purposes since they are fast and not prone to oscillation (like Nose-hoover). V-rescale can also be used for production, unless your problem absolutely hinges on how the kinetic energy is handled. Nose-hoover is good for production MD but, due to its oscillation issues, it is not meant for equillibration purposes. Plus, it can be 4-5 times slower than V-rescale.

Furthermore, before planning you simulation, you should first read up on your force field's literature and its Gromacs implementation on what thermostats/barostats are more applicable than others. Some force fields work better with specific algorithms because they have been parametrized with these algorithms in mind.

It all boils down to what you want to study. If you want to investigate values that are highly dependent upon the correct velocity/temperature distribution (like heat capacity), then you want a strict algorithm for your production run. If you don't care so much about that, either solution will do (in protein simulations we rarely have any issues of the sort). If your simulation runs with the two different couplings converge to a common (not identical, but in the same area of confidence intervals) result for your phenomenon, you're fine with either thermostat option.

The above idea applies, to certain extents, to barostats as well.

Thank you, Fotis for detailed answer. Of course I would run wet experiments or simulations more than once before I believe the results. I will keep these things in mind.