[gmx-users] Re: ewald_rtol and convergence

Michael Shirts mrshirts at gmail.com
Wed Feb 22 19:02:45 CET 2006


So, it turns out that the problem with the poor convergence is that I
accidentally had pbc = no set instead of pbx = xyz in the mdp.  Once
that was fixed, covergence was much more reasonable.

I'm wondering if there should be a warning in the code that says "Do
you REALLY want to use Ewald with a non-periodic system?"

Thanks,
Michael

On 2/17/06, Michael Shirts <mrshirts at gmail.com> wrote:
> Hi, all-
>
> So, I'm wondering what the effect of ewald_rtol is.  My understanding
> is that it should just be shifting the calculation between the real
> space and the reciprocal space summations.  But the total
> electrostatic is changing non-negligibly as well.  My system is a box
> of waters with a protein -- total charge is zero.  PME Cutoff is 0.9,
> fourier_spacing is 0.02 (so well converged in all cases), PME order 6.
> Octahedral box, (5.78643,5.45550,4.72461) box vector, all the same
> configuration.
>
> Ewald_rtol                   total                  Short range
>  Long range         1/beta (nm)
> 1e-02                  -53195.725708   -46946.948640     -6248.777068
>      0.494129
> 1e-03                  -53432.220675   -45014.961538     -8417.259138
>      0.386805
> 1e-04                  -53652.773007   -43093.541701   -10559.231306
>     0.327146
> 1e-05                  -53872.777698   -41119.255332   -12753.522366
>     0.288146
> 1e-06                  -54092.976378   -39070.390423   -15022.585955
>     0.260198
>
> I did some comparison to straight Ewald, as well.
>
> Ewald_rtol                   total                  Short range
>  Long range
> 1e-02                   -52899.579997   -46946.948640    -5952.631357
> 1e-03                   -53001.090875   -45014.961538    -7986.129337
> 1e-05                   -53182.993892   -41119.255332   -12063.738560
>
> We see that the short range part is exactly the same in PME and
> straight Ewald, as it should, though the long range part is not the
> same, even when the PME part is well converged (i.e., I reduced the
> fourier_spacing until any changes were sub 0.002 kJ/mol, which is one
> part in 25 million).
>
> I'm unsure why the differences.  It there coding issue with the
> octahedral box? Is there some effect with the dielectic at infinity
> (i'm using epsilon_surface = 0).  What is the 'true' electrostatic
> energy?
> I tried different cutoffs -- presumably, as we take a low rtol,
> pushing the calculation further into the direct space, and choose a
> longer cutoff, the correction will get more accurate.
>
> PME: (order 6, ewald_rtol = 1e-02)
> cutoff                     total                  Short range
> Long range             1/beta
>
> 0.9                  -53195.725708   -46946.948640     -6248.777068
>    0.494129
> 1.8                  -52875.688467   -49910.307947     -2965.380520
>    0.988258
> 2.7                  -52798.032604   -50836.767614     -1961.264991
>    1.48239
> 3.6                  -52764.124490   -51296.887618     -1467.236872
>    1.97652
> 4.5                  -52745.340510   -51572.825196     -1172.515314
>    2.47065
>
> Ewald: (fourier_nx,ny,nz = 25)
>
> 0.9                  -52899.579997   -46946.948640    -5952.631357
>     0.494129
> 1.8                  -52785.799065   -49910.307947    -2875.491118
>     0.988258
> 2.7                  -52761.626727   -50836.767614    -1924.859113
>     1.48239
> 3.6                  -52747.927053   -51296.887618    -1451.039435
>     1.97652
> 4.5                  -52738.223387   -51572.825196    -1165.398191
>     2.47065
>
> So it seems that -52730 is closer to the "right" answer.  Why isn't
> this more possible to with other combinations of terms?  Am I missing
> something?  Is this level of agreement (both for Ewald, and between
> PME and Ewald) an unreasonable expectation? :)
>
> Thanks,
> Michael
>



More information about the gromacs.org_gmx-users mailing list