Using TCE with Douglas-Kroll Hamiltonian (fock or nofock)

[Stephen H. Yuwono]

Dear NWChem devs,

I am trying to run some coupled-cluster calculation using TCE and I would like to run it with (scalar) relativistic Hamiltonian, using the 3rd-order Douglas-Kroll (DK) approach.

I read in the documentation for TCE that I should be using "nofock" when using a DK reference. I tested it using NWChem 7.0.0 on the O2 molecule, employing CCSDT/cc-pVTZ-DK and freezing the 1s orbitals, and I notice that while the SCF energies are identical (to within the convergence threshold), the CCSDT energies differ by ~40 millihartree. The outputs corresponding to these calculations are attached (O2_CCSDT_TZ-DK_*.log) and as you can see, they are identical except for the "nofock" keyword.

Interestingly, I tested the same on the Mg2 dimer, using CCSDT/aug-cc-pV(T+d)Z-DK and correlating only the 3s orbitals, but the runs with and without "nofock" do not produce any difference. Both the total SCF and CCSDT energies are identical between the two runs. I also tested with CCSDT/aug-cc-pwCVTZ and freezing the 1s orbitals, with the same results. The outputs are also attached for your convenience (Mg2_ATdZ-DK_CCSDT_*.log and Mg2_AwCTZ-DK_CCSDT_*.log).

Is the non-difference happening in Mg2 just a coincidence or is there a mistake in the calculation that I introduced? Should I still use "nofock" in these DK-based calculations? I would really appreciate your advice on this issue.

As always, thank you very much for your help!

Best regards,

Stephen

[Stephen H. Yuwono]

Dear devs,

Upon a closer inspection, it looks like the difference between "fock" and "nofock" in my O2 calculation is because O2 is an open-shell system. I tested this by running non-relativistic calculations for the O2 at the frozen-core CCSDT/cc-pVTZ level of theory. Again, the total SCF energy is the same but there is a ~40 millihartree difference in the correlation energies between the runs with and without the "nofock" keyword (please see the attached O2_TZ_CCSDT_*.log files). I ran another closed-shell system, HF, using CCSDT/cc-pVTZ-DK with DK3 and, as in the case of Mg2, I found no difference in the correlation energy (cf. the attached hf_*.log files).

Thus, in addition to the question about using "nofock" when running a relativistic calculation, I have another one: Should we recompute the Fock matrix for open-shell calculations using TCE? Also, what about the behavior of the non-TCE CCSD module when using DK hamiltonian or computing open-shell systems?

Thank you in advance!

Best regards,

Stephen

[Edoardo Aprà]

The NOFOCK trick does not work with ROHF (we should probably have NWChem bail out when ROHF & NOFOCK are present ...)

Replacing ROHF with UHF should make FOCK and NOFOCK energies agree

[Stephen H. Yuwono]

Hi Edoardo,

I see, thank you for the explanation. I will test some runs using UHF reference.

If I want to use ROHF reference, does this mean that the default FOCK behavior is the appropriate choice in both open-shell and relativistic calculations? I just want to confirm before committing to a production run for such calculations.

Thank you!

Stephen