[gmx-users] Re: Problems with cg

[Erik Lindahl]

Hi,

tccfr at ig.com.br wrote:

> Dear Erik, 
> First of all I would like to thank you for your help in my last mail. It 
>was very usefull. 
> But now I am facing another problem, this time with minimization using cg. 
>I have minimized one protein (DHFR) with steep until 201 Kj/mol.nm and now, 
>I am trying to go on with cg. In my first round, I fixed a emtol of 200 just 
>for testing. GROMACS stopped in step 1890 and output the following mesage: 
>"Negative w: -2.828724731694e+17 
>         gpa = 2.656058552315e+08  gpb = 1,101866243368e+09 
>         a = 0,000000000000e+00, b = 1,404184966477e-08 
>         EpotA = -7.676745000000e+05, 
>EpotB = -7.676685625000e+05  Negative number for sqrt encountered 
>(-282872473169428480000000) 
>        Terminating minimization 
>      Conjugate gradients did not converge in 10000 steps 
>      Function value at minimum = -7.6767e+05" 
>
>This output put on me the questions: 
>
>1) The value of Fmax at the moment of termination was 3.10116e+03. This 
>value is smaller than the starting Fmax (201). How could it be possible ? Is 
>that (3.10116e+03)a local minimum ? If so, how can I avoid it ? 
>
We're actually minimizing the potential energy - the Fmax is just 
information about the largest force on any atom in this conformation.

>2) What is w ? 
>3) How can I avoid that w become negative ? or what should I do when this 
>happen ? 
>
You'll probably have to try double precision, or set nstcgsteep to e.g. 
100 (perform a steepest descent step every 100 cg steps); there are a 
couple of accuracy-sensitive operations in conjugate gradients, so 
sometimes single isn't enough.

Cheers,

Erik