[molpro-user] Mixed derivatives in Numerical Hessian
Werner Győrffy
gyorffy at theochem.uni-stuttgart.de
Wed Feb 8 00:24:09 CET 2017
Dear Aleksandr,
I agree that the thresholds must be made tighter: especially for
Hartree-Fock, and also for CP-HF in the case of using analytical
gradients. One should be careful not to set the thresholds too tight.
For example, the suggested "gthresh,orbital=1.0d-10" is equivalent to
"{hf;accu,20;}" which seems to be unnecessary low and might cause
convergence problems.
Yes, MP2 and DF-MP2 Hessians are computed by using analytical first
derivatives.
CCSD(T) analytical gradients have been already implemented in the
development version of Molpro. It will be hopefully available soon in
the next release.
Regards,
Werner.
On 02/06/2017 07:14 PM, sjk wrote:
> I find that tightening the convergence criterion on various evaluations
> generally solves the low frequency hessian convergence problems (my
> experience is that anything below about 200 cm-1 should not be
> considered reliable with the standard parameters).
> Thus, I generally insert the following line at the top of my input file
> gthresh,energy=1.0d-10, orbital=1.0d-10, oneint=1.0d-16, twoint=1.0d-16,
> optgrad=1.0d-6, compress=1.0d-13
> Best Regards,
> Stephen
> On Feb 6, 2017, at 9:12 AM, Leonid Shirkov <leonid.shirkov at gmail.com
> <mailto:leonid.shirkov at gmail.com>> wrote:
>
>> Dear Colleagues,
>>
>> in some specific cases, the current accuracy of the numerical hessians
>> is not enough, e.g. for very low frequency torsional vibrations (~50cm-1).
>> The CCSD(T) freq analysis gives imaginary values instead of the real ones
>> for the lowest modes. The solution is to find manually the hessian in
>> the internal coordinates
>> and then find the eigenvalues of the GF matrix, but that is a lot of work.
>>
>> If MP2 is used for such cases, then there are no imaginary frequencies.
>> Do I understand correctly, that for MP2 freq analysis the hessians are
>> found
>> by differentiating the analytical MP2 gradients?
>>
>> Using the analytical gradients for highly accurate methods like
>> CCSD(T) would probably resolve the problem,
>> but they are not currently available in Molpro.
>>
>> Best regards,
>> Leonid
>>
>> On Mon, Feb 6, 2017 at 10:18 AM, Werner Győrffy
>> <gyorffy at theochem.uni-stuttgart.de
>> <mailto:gyorffy at theochem.uni-stuttgart.de>> wrote:
>>> Dear Aleksandr,
>>>
>>> Numerical Hessians in Molpro are computed by using central finite
>>> differences with a 2-point formula as a default. That is a "general
>>> formula". That gives accurate results in most of the cases. There is a
>>> trade-off between accuracy and efficiency: More accurate finite field
>>> calculations would increase the number of single point calculations
>>> significantly. If one needs more accurate Hessians, it can be done
>>> only by
>>> computing that manually by using procedures.
>>>
>>> Regards,
>>>
>>> Werner.
>>>
>>>
>>> On 02/04/2017 02:04 AM, Aleksandr Lykhin wrote:
>>>>
>>>> Does anybody know how Molpro calculates mixed derivatives using central
>>>> differences? It seems like it generates only two mixed displacements
>>>> instead of four, so the general formula cannot be applied directly.
>>>>
>>>> --
>>>> Kind regards, Aleksandr O. Lykhin.
>>>>
>>>>
>>>> _______________________________________________
>>>> Molpro-user mailing list
>>>> Molpro-user at molpro.net <mailto:Molpro-user at molpro.net>
>>>> http://www.molpro.net/mailman/listinfo/molpro-user
>>>>
>>> _______________________________________________
>>> Molpro-user mailing list
>>> Molpro-user at molpro.net <mailto:Molpro-user at molpro.net>
>>> http://www.molpro.net/mailman/listinfo/molpro-user
>> _______________________________________________
>> Molpro-user mailing list
>> Molpro-user at molpro.net <mailto:Molpro-user at molpro.net>
>> http://www.molpro.net/mailman/listinfo/molpro-user
>
>
>
> _______________________________________________
> Molpro-user mailing list
> Molpro-user at molpro.net
> http://www.molpro.net/mailman/listinfo/molpro-user
>
More information about the Molpro-user
mailing list