Re: ABF thermostats and energy introduced into the system

From: Jérôme Hénin (
Date: Mon Feb 25 2013 - 08:01:52 CST

Dear Ajasja,

----- Original Message -----
> Dear all,
> This next questions is similar to asking "I poured kerosene into my
> car -- why did it explode?".

That's because you should have poured the kerosene into the gas tank only.

> I performed 400 ns of ABF simulations on the phi and psi angles of
> the alanine dipeptide in vacuum.
> Then I made a short run without using the thermostat. This was done
> on purpose and I did of course notice the warning printed by the
> colvars module. The net result of this was that my system heated up
> considerably.
> I realize that meta-dynamics always puts energy into the system, but
> I had the feeling that since ABF is a time dependant potential it
> could input energy in some bins and output energy in other, so that
> the net work done on the system would be small. So my questions are:
> * Does ABF always pump energy into a system and what factors
> determine if energy goes in or out?

It's an interesting question for which I don't have a well-argued answer. My intuition is that this is linked with starting from a low energy conformation, exploring higher energy conformations "for free" (using the generous ABF uphill bias, which keeps no memory that the system "owes it" some energy), then sliding down the same slope back into the initial low-energy basin and picking up momentum, as this time around the ABF bias is almost time-independent. You could try starting from a high-energy conformation and see what happens.

> * What thermostat is the most suitable for ABF? (Langevin,
> tCouple, Lowe-Andersen, ...) Probably the best thermostat is the
> one that does not hinder diffusion too much, so the answer would
> be Lowe-Andersen, but that does not work with CUDA, so the next
> best is Langevin?

>From a statistical perspective, any "canonical" (as in canonical ensemble) thermostat is fine. But the real bottom line is, I see no excuse for anyone to use tCouple under any circumstances. I know, I've been saying that for a long time, now get off my lawn (-gevin). Lowe-Andersen is great, but not having with CUDA is not a show-stopper unless you rely on particularly sensitive kinetic measurements, or simulate a system with mesoscopic flow, or far from equilibrium etc. In an average-sized system, you can tune Langevin to introduce no noticeable friction while maintaining the temperature within acceptable bounds.


This archive was generated by hypermail 2.1.6 : Wed Dec 31 2014 - 23:20:57 CST