Re: [AMBER] DIfferences between thermostats

From: Marc van der Kamp <>
Date: Wed, 23 Nov 2016 15:53:48 +0000

Thanks for sharing your observation about Berendsen with long timescale
simulation, Dave Case!
I too was thinking I could use a weak Berdensen thermostat just to keep the
overall energy
from drifting, i.e. to get an NVE-like behaviour.
I have not yet seen problems, but I have only done (multiple) 500ns
simulations so far.

Perhaps I need to switch to Langevin with very low collision constant...


On 23 November 2016 at 01:36, David A Case <> wrote:

> On Tue, Nov 22, 2016, Dave Cerutti wrote:
> > Yes, fair enough--Berendsen is bad, I'm with Bernie (Brooks)! A very
> weak
> > Berendsen thermostat is what was recommended to me for simulations where
> > diffusion properties and NVT are needed simultaneously, by one of our
> > respected colleagues no less. And, I agree--formally, it's bad, but with
> > weak thermocoupling, in practice the effects are negligible.
> In the bad old days, we used Berdensen and got "flying blocks of ice". We
> found that by tightening up tolerances one could more or less get rid of
> this. But that was for runs of dozens of nanoseconds. I recently thought
> I could use a weak Berdensen thermostat just to keep the overall energy
> from drifting. Seemed to work great for about 8 microseconds, at which
> point the protein froze--just like the old days, but now after 100 times
> longer simulation. Rules of thumb that seemed good for simulations of
> 10^7 steps may fail for runs of 10^10 steps.
> I suspect (following work from Leimkuhler, Izaguirre, Skeel, etc.) that
> Langevin with a really good random number generator and a very low
> collision
> constant would have been a better choice for me. But I don't have personal
> experience to back that up.
> ...dac
> _______________________________________________
> AMBER mailing list
AMBER mailing list
Received on Wed Nov 23 2016 - 08:00:04 PST
Custom Search