reference wavefunction for UCCSD(t)

[Pamela Araneda]

Dear Users,

I am trying to optimize a transition state at UCCSD(t)-F12/cc-pvtz the reference function is RHF but gradients failed to converge for RHF However at UHF it is converging properly. Is there any way that UCCSD(t)-F12 optimization reads the orbitals of UHF instead of RHF.

Thank you,

Best,

Pamela

[tibo...@gmail.com]

Dear Pamela,

Could you please specify exactly what you mean by "gradients failed to converge for RHF"? Are you using numerical gradients and RHF (which stands for ROHF in this context) is failing to converge at one of the displacements? Or are you seeing a convergence failure of the CC iterations in UCCSD(T)-F12?

There are a number of things you may try to tweak to achieve convergence, before you go for the fairly big hammer of using a different reference wavefunction for CC. For example if you have issues in the RHF part, you could try RHF, SO-SCI or RHF, SO to use different HF solver algorithms, which can often help.

Best,

Tibor

[tibo...@gmail.com]

Dear Pamela,

When running the MCSCF program (MULTI), it is often necessary to explicitly specify both the number of closed and occupied orbitals.

If you want to do doublet ROHF with the MCSCF program like this, you should always see 1 electron in 1 active orbital, and 1 determinants at the beginning of MULTI's output.

Also please note the difference in syntax between options and directives, the SO-SCI option got silently ignored in your input because you used it as an option for the closed directive. You can tell it was ignored, because it printed "Second-order MCSCF: L-BFGS accelerated" instead of "First-order MCSCF: SO-SCI".

Although in this particular case it has very little consequence, options getting ignored can lead to a lot of frustration.

But this input would likely not run even after you correct this issue, as the MCSCF program outputs natural orbitals, which the CC program cannot use directly, as the error message is telling you.

A workaround is to insert an RHF calculation after the MCSCF, which loads the natural orbitals as the guess. It should instantly converge and not change the energy compared to the MCSCF before it, but it will write canonical orbitals to disk, which should then get picked up by the CC program.

The following should work, but please double check it yourself as well.

It is easy to accidentally get wrong energies if one makes a mistake in the active space or wavefunction specification.

The record and file number chosen does not have too much significance.

{MULTI, SO-SCI; closed,24; occ, 25; orbital,1500.3}    !saves the final natural orbitals to record 1500 of file 3
{RHF; start,1500.3}    !starts the ROHF from the orbitals in record 1500 of file 3
{UCCSD(T)-F12b}

Best,

Tibor

On Tue, Jul 2, 2024 at 12:11 AM Pamela Araneda <pamela.a...@gmail.com> wrote:

Hi Tibor,

Thank you so much for explaining the details.

I tried attempting MULTI, SO-SCI but facing memory issues! File attached. Is the input correct or am I missing something?

Best,

Pamela

On Sun, Jun 30, 2024 at 2:44 PM Tibor Győri <tibo...@gmail.com> wrote:

Dear Pamela,

The issue is indeed that your ROHF is not converging.

I have noticed that you are using a somewhat old version of Molpro, 2020.1. I recommend that you look into upgrading to the latest version, because newer versions have a few features that make dealing with HF convergence issues easier, especially for new users. See the documentation (https://www.molpro.net/manual/doku.php?id=the_scf_program) for more details.

I have looked at your input and found that it had multiple issues. First, optg was put before the CC computation so you would have likely done an optimization at the HF level, followed by a CC single-point energy at the final geometry. This is likely not what you wanted to do. A second problem was your choice of basis set, cc-pVDZ. Using this with F12 methods is generally not advised as the F12 correlation correction benefits a lot from having some diffuse basis functions in the basis set. You should use either aug-cc-pVDZ or cc-pVDZ-F12 instead of cc-pVDZ. These are larger and make the computation slower but if you want good results from CCSD(T)-F12, then cc-pVDZ is probably not good enough. Perhaps you can consider doing some preliminary optimizations using DFT, and then refining your geometries with a more expensive method.

I have attached the output of me trying to run a modified version of your input file using Molpro 2023.2. As you can see, with SO-SCI, the ROHF/VDZ-F12 computation at your starting geometry converges quite smoothly, the subsequent CC iterations also succeed, and the geometry optimization would start after (T) finishes, which is something I did not wait for. If you cannot upgrade your Molpro version, this input will probably not work as-is, but you should be able to achieve the same results with MULTI,SO-SCI.

Best,

Tibor Győri

On Sun, Jun 30, 2024 at 3:32 AM Pamela Araneda <pamela.a...@gmail.com> wrote:

Hi Tibor, 

Thanks for your reply. For more clarity of the issue the files are attached herewith. Actually I am trying to optimize transition state geometry at UCCSD(t)-F12/cc-pvdz for which it uses a ROHF reference function that is not converging.

Looking forward to your response,

I am using MOLPRO for the first time. A little more details can really help me out.

Best,

Pamela

--
You received this message because you are subscribed to the Google Groups "molpro-user" group.
To unsubscribe from this group and stop receiving emails from it, send an email to molpro-user+unsubscribe@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/molpro-user/eaa44ea7-8b6b-4b2b-957f-aacf701f29e6n%40googlegroups.com.