Seeking suggestions on running MRCI using RAS3 code

[ankit7540]

Dear all,
This question goes in tandem with my previous questions on MCSCF/RASSCF (viewtopic.php?f=9&t=1074) on this forum.
I am trying to run calculation for simple diatomics using RASSCF aiming to get polarizability. Tests with basis sets have shown slow change in alpha values and augQZ looks feasible. However, convergence of alpha values over number of e- in RAS3 and the size of RAS3 has been slow. This accompanies the slow convergence of MCSCF calculations.

For example,

N2 (File is attached)

Code: Select all

*CONFIGURATION INPUT
.SYMMETRY
1
.SPIN MULTIPLICITY
1
.RAS1 SPACE
1 0 0 0 1 0 0 0
.RAS2 SPACE
1 1 1 0 1 1 1 0
.RAS3 SPACE
4 2 2 1 3 2 2 1
.INACTIVE ORBITALS
1 0 0 0 1 0 0 0
.ELECTRONS
10
.RAS1 HOLES
0 2
.RAS3 ELECTRONS
0 2

gives,

alpha_xx: 10.03033613
alpha_zz: 14.54827311 in atomic units

These values are lower than CCSDT results at the same internuclear distance (10.2351 and 14.8425 a.u.). More tests showed that running with a larger basis, including Rydberg orbitals and also all 14e- as active does not account for this difference in the values of alpha.

Increasing, number of e- to RAS3 increases alpha (10.08330, 14.6112) while increasing size of RAS3 decreases the values (9.9500, 14.36988). (symmetry is broken in the large RAS3 case, with RAS3 set to 63315331)

By performing similar sized CASSCF calculation on MolPro, it is found that there are several low lying states, which are not included when opting for RASSCF with a small RAS3. This is a reason for increasing the size of RAS3. But, then large RAS3 is slow to optimize, which has been explained earlier on my questions.

So, this question on a different approach. Can I do this.

1. Run CASSCF and keep the SIRIUS.RST file
2. Setup my RASSCF calculation as : Keeping RAS2 same as previous CASSCF, 2e- in RAS3 and rest all orbitals to virtual, and simulate something like MRCI. Pass SIRIUS.RST as orbital input.

3. On a side note, can we just do CI calculation and not to orbital optimization ( I way I think is to set the Macro iterations in *OPTIMIZATION to 0).

Thanks.

[taylor]

You can certainly do CI calculations rather than MCSCF: this is described in the manual. And using your proposed approach of doing MCSCF, saving the orbitals, and then using them as input into a CI calculation will work.

That said, it is not clear to me that this is a fruitful approach, unless you are only interested in N₂! One problem is that your approach will not be size-extensive. This would make comparisons between systems with different numbers of correlated electrons difficult. Another relates to your using CCSDT as your comparison standard. The code I use for full CCSDT is CFOUR, and by default this correlates all electrons. This is not necessarily a disadvantage provided the basis set is capable of describing core and core-valence correlation. This would mean something like aug-cc-pCVnZ, for example (aug in this case because you want polarizabilities). Or, of course you can freeze the cores in the CCSDT calculation. Given that you have not correlated the cores in your Dalton calculation it is important to eliminate any issues from this source. Finally, it might be worth comparing the full CCSDT results with CCSD(T) results. If these two methods give very similar results, it would suggest that higher-excitation (that is, in CI language, multiconfigurational) issues are not the problem. If CCSD(T) and CCSDT differ significantly, then not only might there be such problems, but also CCSDT itself might not be adequate.

Best regards
Pete

[hjaaj]

You can indeed do RASCI (basically replace .MCSCF with .CI plus keywords to read in MCSCF orbitals and CI coefficients.
If you want to enlarge the CI space compared to the RASSCF by including more RAS3 orbitals, that will not be particularly good, because the added orbitals are not optimized for correlation. However, it makes a lot of sense to extend to e.g. a CI-SDT or CI-SDTQ on top of an MCSCF-SD.

That said, you will get poor polarizabilities from the **RESPONSE. It will perform a "CI response" which corresponds to a sum-over-states expression over all CI eigenfunctions (without calculating the eigenfunctions, that is the smart part of it). You will get poor values for the polarizabilities because the occupied space optimized for correlation of the ground state will not contain the diffuse orbitals needed for good polarizability values.

[ankit7540]

Thank you for the comments.

Size consistency of the proposed CI (or even RASSCF) approach is a concern. I need polarizabilities over a wide internuclear distance and using rovibrational wavefunctions, I want to compute the relevant transition matrix elements. I have corresponding experimental data for the same.

Looking towards coupled cluster approach, I wonder if CFour package allows for CC calculations with open-shell systems. Since, in my case I have two molecules to be studied, N2 and O2 and for O2 I need proper treatment for the ground triplet state.
From the online page of cfour (http://www.cfour.de/) it appear that this should be possible.

[taylor]

This is of course the Dalton forum, but certainly CFOUR can do open-shell coupled-cluster calculations, which Dalton cannot. (Dalton can do triplet excited states using a closed-shell coupled-cluster reference function, but O₂ is awkward because neither of the low-lying singlets is a closed-shell state.)

Size-consistency is not enough. As an example, consider the noble gas atoms, where the question of size-consistency does not arise because there is no dissociation process: we are looking at a single atom. If you wish to compare the effects of core correlation on the properties as you go from He to Xe, you need to use a method that is size-extensive (scales correctly with the number of particles). In any event, whichever extensivity/consistency property you consider, the sort of CI you propose will not have it.

Best regards
Pete