Geometry Optimisation of Large Charged Systems

[Olivia Lynes]

Hi all,

I'm having a problem with a few of my geometry optimisations not converging after several thousand steps and having looked at the energy changes through the main output file it's not decreasing but is fluctuating a lot in a certain range. I'm at a loss to why as I've got similar calculations completed with similar input and just a different box size and they have converged without this fluctuation in a few hundred steps.

I've attached the input, coordination files and the first hundred or so steps of the output file.

These are geometry optimisations of an Mg 2+ ion in a box of 64 waters, with a cell size of 13.41 angstroms, so the overall system is charged 2. I've used the GTH DZVP basis sets and potentials. 

Having done a cell optimisation on just the 64 waters, we introduced the Mg ion and charged the system. In lieu of being able to do cell optimisations on a charged system, then geometry optimisations are being done at varying box sizes by 10ths of an angstrom as a sort of manual cell optimisation to try to find both a minimum energy structure and a minimum energy box size. 

I've done similar calculations using Ca and Sr and haven't had this issue with any of my jobs, they all converge in a few hundred steps. 

Is this potentially a problem using charged systems? Is it necessary for me to put a counter ion in which would let me do a standard cell optimisation? 

Or is this approach completely wrong and we would be better off doing MD? 

The aim is to do ab initio MD on these systems to look at the coordination of the ion in a bulk of water, so reaching a minimum energy structure and box size isn't essential but it would be nice to have the regular DFT data to back it up. 

Thanks in advance for any insights.

Olivia

[Marcella Iannuzzi]

Dear Olivia,

The SCF in your run is never converging. Forces  calculated from a not converged SCF are affected by large errors,  which explains the behaviour of the geometry optimisation. 

One shouldn't  carry on a geometry optimisation run when the SCF does not converge.

Allowing a larger number of SCF iterations might help (see OUTER_SCF / MAX_SCF).

Kind regards

Marcella 

[Rolf David]

Hi Olivia & Marcella, 

I allow myself to jump back on this problem/answer, and ask the CP2K team/users:

For a MD calculation (or maybe other, geometry for example here), is there a way to stop it if the SCF hasn't converged (after the inner/outer SCF) ? 

I try to find in the manual but no infos (or i looked at the wrong place ?)

If one SCF didn't converge, forces are still calculated and MD goes on to next step. So I kill it manually, but making it automatically save CPU usage and tell you right away: "No convergence, please check".

Thanks

Best regards,

Rolf

[Matt W]

Hi Olivia,

Mg2+ is nearly pathological if you use the default q10 pseudo and the GPW method. It needs a plane wave cutoff of around 1200 Ry or so to start getting reasonable answers. If you are sticking with non-hybrid functionals, like PBE, then try using the GAPW method (QS/METHOD section) and see if that helps. A cutoff of around 350 Ry should give well converged numbers then.

Matt

[S Ling]

Hi

Following what Matt has said, i.e. you might need to use an extremely large cutoff for the q10 pseudoptential of Mg, if you are happy with the GTH-PBE-q2 pseudopotential of Mg with only two valence electrons for your system, you may try the attached DZVP-q2 basis set which I have reoptimised for Mg to be used with the GTH-PBE-q2 pseudopotential. You should be able to get reasonable results with a normal cutoff (e.g. ~400 Ry) using the q2 pseudopotential of Mg.

SL

--
You received this message because you are subscribed to the Google Groups "cp2k" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cp2k+uns...@googlegroups.com.
To post to this group, send email to cp...@googlegroups.com.
Visit this group at http://groups.google.com/group/cp2k.
For more options, visit https://groups.google.com/d/optout.

[Olivia Lynes]

Hi Matt,

Thanks for that info. Is this just a problem that's specific to Mg? I'll see what happens with the GAPW method, do you know if it's much more expensive than the standard GPW method?

Thanks

Olivia

[Olivia Lynes]

Many thanks for the pseudopotential! It may be a bit small, but we shall see.

Olivia

[Marcella Iannuzzi]

Hi Olivia, 

GAPW should not be much more expensive, in particular if  a reduced cutoff can be used.

Coming back to your original e-mail. Indeed, the optimisation of the geometry of a liquid water box makes little sense, in my opinion.

The internal pressure of liquid water is obtained from the fluctuations at finite temperature. A frozen configuration is close to useless.

Finally, to answer the question of Rolf: no, there is no automatic run-stop if SCF does not converge. 

In most of the cases, one does neither want nor need such a stop. It is true that for pathological systems it could prevent wasting of time. 

regards,

Marcella