[molpro-user] OQVCCD does not converge at extreme geometries, but CCSD does
Tibor Győri
tiborgyri at gmail.com
Thu May 25 17:48:24 CEST 2017
Dear Daniel,
Thank you very much for your multiple suggestions, and taking the time to
also run them. I had no idea it is possible (or indeed advisable) to use
the amplitudes of a different kind of coupled cluster method, and I don't
think the nomp2 option for OQVCCD is documented in the Molpro manual.
Anyways, I will gather the geometries with QVCCD convergence issues (less
than 1000 out of ~50k, some have HF convergence problems and other weirder
failures) and try the zero-vector guess first, since it saves the time
required by the BDCD calculation.
Best wishes,
Tibor
2017-05-23 12:25 GMT+02:00 Daniel Kats <dnkats at gmail.com>:
> Dear Tibor,
>
> There are few things one can try:
>
> 1. The best solution would be to go from a single-reference geometry to
> the multi-reference geometries in small steps, always restarting from the
> previous calculation, e.g.,
> file,2,molecule.wfu
> .....
> hf
> {oqvccd(t),shiftp=0.8,shifts=0.6;maxit,100;start,5200.2;save,5200.2}
>
> 2. If it's not possible, one can try to generate a more reasonable
> starting guess than MP2.
> First option is to simply use a zero-vector instead of MP2-amplitudes:
> {oqvccd(t),shiftp=0.8,shifts=0.6,nomp2=1;maxit,100}
>
> 3. Another option is to use e.g. distinguishable cluster amplitudes as a
> starting guess for OQVCCD:
> {bdcd,nomp2=1,shiftp=0.8,shifts=0.8;maxit,100;save,5200.2}
> {oqvccd(t);maxit,100;start,5200.2}
>
> You can find outputs of these last two options attached.
>
> Best wishes,
> Daniel
>
> On Tue, May 23, 2017 at 9:01 AM Tibor Győri <tiborgyri at gmail.com> wrote:
>
>> Yes, I have tried {oqvccd(t);shift,1,1} and {oqvccd(t),shifts=0.2,shiftp=0
>> .2,thrdis=1.0;diis,1,1,15,1;maxit,100;}
>> Since I only tried these at a single geometry, it is possible that they
>> would help at the other problematic geometries, but they are no silver
>> bullet.
>>
>> 2017-05-22 7:42 GMT+02:00 Peterson, Kirk <kipeters at wsu.edu>:
>>
>>> Dear Tibor,
>>>
>>> so you tried shifting both the singles and pairs via something like:
>>> {oqvccd(t);shift,1,1} ?
>>>
>>> best,
>>>
>>> -Kirk
>>>
>>> > On May 21, 2017, at 5:54 AM, Tibor Győri <tiborgyri at gmail.com> wrote:
>>> >
>>> > Dear Molpro users,
>>> >
>>> > Recently I have run into some issues while trying to use OQVCCD(T) for
>>> the purpose of calculating energies for PES fitting. The problem is, that
>>> to fit a global PES some calculations have to be run at pretty extreme
>>> geometries, and I have seen many examples, where ordinary CCSD and CCSD-F12
>>> converges but OQVCCD does not.
>>> >
>>> > The awesome multiple bond breaking characteristics of OQVCCD(T) would
>>> be really advantageous for fitting a global PES, but these convergence
>>> issues are hindering my efforts. I have tried to increase the maximum
>>> number of iterations allowed to 100, but a number of calculations still do
>>> not converge.
>>> > I also tried to use level shifts and increased DIIS size at one
>>> geometry, but it did not work any better.
>>> >
>>> > Is there anything I could do to increase the chance of OQVCCD
>>> converging in difficult cases? Since these input files are automatically
>>> generated, anything that requires manual per-input tuning is pretty much
>>> out of the question.
>>> > I have attached an abridged log file of a calculation that had issues.
>>> >
>>> > Best wishes,
>>> > Tibor Győri
>>> > MSc student
>>> > University of Szeged
>>> >
>>> > <oqvccd.log>_______________________________________________
>>> > Molpro-user mailing list
>>> > Molpro-user at molpro.net
>>> > https://urldefense.proofpoint.com/v2/url?u=http-3A__www.molp
>>> ro.net_mailman_listinfo_molpro-2Duser&d=DwICAg&c=C3yme8gMkxg
>>> _ihJNXS06ZyWk4EJm8LdrrvxQb-Je7sw&r=mqaXWsZ0ghQR-pXvPLjY_g&m=
>>> OtWbWWEZBfJvvNsgh34ZzXKGiVsLFXfovJUhURi27qM&s=M6yVxvlkDsAB_0
>>> 9PjJGiboroHxeAxsy3-OuOSeigYfk&e=
>>>
>>>
>> _______________________________________________
>> Molpro-user mailing list
>> Molpro-user at molpro.net
>> http://www.molpro.net/mailman/listinfo/molpro-user
>
> --
> Daniel Kats
> Institute for Theoretical Chemistry
> University of Stuttgart
> Pfaffenwaldring 55
> D-70569 Stuttgart
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.molpro.net/pipermail/molpro-user/attachments/20170525/beb7f47d/attachment.html>
More information about the Molpro-user
mailing list