Crash of DF-CCSD or DF-CCSD(T) geometry optimization (MRCC package)

Question

I am interested in geometry optimization at the CCSD(T) or DF-CCSD(T) level, with MRCC. The calculation, however, crashes when executing prop:

 ************************ 2022-11-30 15:42:32 *************************
 Executing prop...

 Allocation of 500.0 Mbytes of memory...

 Number of integral batches:                      1

 Fatal error in exec prop.
 Program will stop.

 ************************ 2022-11-30 15:42:32 *************************

giving the error:

forrtl: severe (24): end-of-file during read, unit 26, file /data/test-opt-mrcc/CCDENSITIES
Image              PC                Routine            Line        Source
prop               0000000005C2DC58  Unknown               Unknown  Unknown
prop               0000000005C6486A  Unknown               Unknown  Unknown
prop               000000000045DB46  Unknown               Unknown  Unknown
prop               0000000000405AD5  Unknown               Unknown  Unknown
prop               0000000000400DF2  Unknown               Unknown  Unknown
prop               0000000005D12DF9  Unknown               Unknown  Unknown
prop               0000000000400CCA  Unknown               Unknown  Unknown

I suspect that it is because of the ccprog=ccsd option, i.e., one can do geometry optimization only when ccprog=mrcc. Is that the case?

A test input:

# CCSD geometry optimization for water with the cc-pVDZ basis set
basis=cc-pVDZ
calc=ccsd(t)
ccprog=ccsd
mem=500MB
gopt=full

unit=bohr
geom
H
O 1 R1
H 2 R1 1 A

R1=2.00000000000
A=104.2458898548

Answer 1

The MRCC manual suggests that analytic gradients are not available for CCSD(T), they are only available for the following 12 categories of methods:

6.2 Geometry optimizations and first-order properties

Available methods

Geometry optimizations and first-order property calculations can be performed using analytic gradient techniques with the following methods. See also Tables LABEL:GrTable and LABEL:ExTable for the availability of analytic gradients for ground- and excited-state models, respectively.

1. conventional and DF (RI) HF-SCF (Ref. 137): RHF and UHF
2. conventional and DF (RI) DFT (Ref. 71): RKS and UKS with LDA, GGA, meta-GGA (depending only on kinetic-energy density), and hybrid functionals as well as dispersion corrections
3. double hybrid density functional methods (Ref. 71), such as B2PLYP, B2PLYP-D3, B2GPPLYP, etc. (current limitations: only MP2 correlation, closed shell RKS, no spin-component scaling, no meta-GGA functionals, only DH functionals for which the DFT contribution to the energy is stationary with respect to the variation of the MO coefficients)
4. DF-MP2 (RI-MP2), currently only for RHF references (Ref. 71)
5. arbitrary single-reference coupled-cluster methods (Refs. 69 and 60): CCSD, CCSDT, CCSDTQ, CCSDTQP, …, CC(n)
6. arbitrary single-reference configuration-interaction methods (Refs. 69 and 60): CIS, CISD, CISDT, CISDTQ, CISDTQP, …, CI(n), …, full CI
7. multi-reference CI approaches (Refs. 70 and 60)
8. multi-reference CC approaches using a state-selective ansatz (Refs. 70 and 60)
9. CIS and TD-HF methods using RHF references
10. arbitrary single-reference linear-response (equation-of-motion, EOM) CC methods (Refs. 60 and 62): LR-CCSD (EOM-CCSD), LR-CCSDT (EOM-CCSDT), LR-CCSDTQ (EOM-CCSDTQ), LR-CCSDTQP (EOM-CCSDTQP), …, LR-CC(n) [EOM-CC(n)]
11. linear-response (equation-of-motion) MRCC schemes (Refs. 60 and 62)
12. The ONIOM approach with arbitrary number of layers [85]. Currently, only the IMOMO [57] part of ONIOM is available in the standalone version of Mrcc with mechanical embedding and automated link atom handling. ONIOM gradients with electronic embedding are also available with perturbation independent point charges.

If you are okay with using numerical gradients, then you can do the geometry optimization manually or via the CFOUR interface:

"Geometry optimizations and first-order property calculations can also be performed via numerical differentiation for all methods available in Mrcc using the Cfour interface."

The upshot of using MRCC is that analytic gradients are available for CCSDT, CCSDTQ, etc., and also for RHF-DF-MP2, but unfortunately not for CCSD(T) or DF-CCSD. CFOUR can do analytic gradients for CCSD(T) but likely not for DF-CCSD or DF-MP2, so with each program there's some unique features, and some missing features.