Dear Psi4 Experts,
I am puzzled by a set of frozen-core DF-CCSD(T) calculations of cyclopropylbenzene that in my hands produce wrong (T) corrections in the aug-cc-pVTZ basis. I believe the correct (T) contribution should be about -0.079 Hartree but I get about -0.125 Hartree with Psi4.
I am running single point energy calculations on MP2/aug-cc-pVTZ geometries for two minima (one of the inputs is attached) as well as for a rotational saddle point. Here is what I see:
1) The (T) correction in CCSD(T)/aug-cc-pVTZ, calculated without density fitting using Dalton is -0.079 Hartree.
2) DF-CCSD(T)/cc-pVTZ gives correct energies for all three stationary points, with (T) corrections about -0.077 Hartree
3) FNO-CCSD(T)/aug-cc-pVTZ with DF-CCSD gives correct energies for all three stationary points, with (T) corrections about -0.079 Hartree.
4) FNO-CCSD(T)/cc-pVQZ with DF-CCSD gives a correct energy for the saddle point, with (T) corrections about -0.083 Hartree; I have not checked the minima with cc-pVQZ. So far, so good
5) DF-CCSD(T)/aug-cc-pVTZ gives incorrect (T) energies for all three stationary points, with (T) corrections about -0.125 Hartree. The effect of (T) on the rotational barrier is now totally wrong (ca 2.5 kcal/mol vs 0.2 kcal/mol). SCF, E2 and CCSD energies are correct with the aug-cc-pVTZ basis.
6) Eliminating one or two of the smallest eigenvalues of the overlap matrix did not help.
7) Tightening various convergence criteria did not help.
8) It did not matter if I used 2, 4, or 6 threads.
9) Both the correct DF-CCSD(T)/cc-pVTZ and the incorrect DF-CCSD(T)/aug-cc-pVTZ calculations chose the LOW_MEM Wabef algorithm in DF-CCSD. All DF-CC calculations are running in C1 symmetry.
10) For a few other (simpler) molecules that I have tested, DF-CCSD(T)/aug-cc-pVTZ calculation seems to give correct (T) energies.
All Psi4 calculations are done with Psi4 1.1a2.dev323, which was compiled from git source with Intel Parallel Studio XE 2017, update 1 (icc 17.0.1 20161005) against MKL with -DMAX_AM_ERI=7. The calculations use Broadwell-E i7-6850K CPU on CentOS.
I am at my wit's end why my DF-CCSD(T)/aug-cc-pVTZ calculation goes bananas for cyclopropylbenzene. Am I overlooking something about how to properly run DF-CCSD(T) calculations? Issues with my compiler or compilation flags? Still some linear dependencies in the overlap matrix? An unexpected feature in the DF-CCSD(T) code? An unexpected behavior of i7-6850K CPU? Any words of wisdom or notes about if this is reproducible on other systems are most welcome.
cpb_avtz_inp.txt (1.2 KB)