[gmx-users] double precision and minimisation error

[Justin Lemkul]

On 11/30/13 9:29 AM, jkrieger at mrc-lmb.cam.ac.uk wrote:
> As has been asked a number of times before, I have seen minimization halt
> at a low number of steps having converged to machine precision when double
> precision is not used - does that not mean that double precision is needed
> at least for minimization? Sorry to ask this question again but I can't
> find a clear answer to this in previous messages.
>

That depends on the outcome.  Consider the purpose of EM - to generate a 
reasonable structure that can be subjected to dynamics, the tolerance for which 
is user-defined in emtol.  Sometimes the input structure requires very little 
change, so you reach an acceptable Fmax (e.g. emtol) within a very small number 
of steps.  Double precision is only required when you are aiming for a very 
small Fmax, as would be necessary for, e.g. NMA calculations.  For most 
dynamics, single precision is perfectly acceptable for all steps in the process.

If EM converges prematurely without achieving the desired Fmax, usually one of 
two things has happened.  Either the input structure cannot be resolved (in 
which case the Epot and Fmax are obviously bad) or the Fmax cannot be achieved 
with the parameters being utilized, in which case *maybe* double precision will 
fare better.  The question you have to ask is whether or not you need that level 
of Fmax.

-Justin

-- 
==================================================

Justin A. Lemkul, Ph.D.
Postdoctoral Fellow

Department of Pharmaceutical Sciences
School of Pharmacy
Health Sciences Facility II, Room 601
University of Maryland, Baltimore
20 Penn St.
Baltimore, MD 21201

jalemkul at outerbanks.umaryland.edu | (410) 706-7441

==================================================