Dear NWChem experts,
I have tried the following input to compute the CCSD(T) energy of a benzaldehyde (C7H6O) molecule.
memory stack 1000 heap 100 global 1000 mb
set lindep:tol 1.0e-06
geometry units angstrom
H 8.75720 7.92160 10.48820
H 8.72660 9.04960 8.06760
O 9.88060 8.29200 6.51940
H 11.81780 4.40680 8.51600
H 10.97580 6.22220 7.12940
H 9.60020 6.10680 11.87500
C 9.41580 7.18320 10.03340
C 9.43300 8.26860 7.70880
C 9.83040 7.20160 8.65980
C 10.67580 6.19220 8.17620
C 11.15960 5.14540 8.97080
H 11.03620 4.34320 11.04200
C 10.74500 5.12720 10.34440
C 9.90000 6.13660 10.82820
end
basis spherical
* library aug-cc-pvdz
end
scf
semidirect memsize 50000000 filesize 0
thresh 1.0e-08
end
ccsd
freeze atomic
thresh 1.0e-06
maxiter 200
diisbas 5
end
task ccsd(t)
The problem here is that during the CCSD iterations the rms error decreases only slowly, and it is hard to get it below 1e-4. Increasing the "diisbas" space to values larger than 5 only helps to stabilize the convergence in general, but still after hundreds of iterations it seems impossible to meet the default convergence criterion for the rms error (less than 1e-6). Playing with the "tol2e" parameter did not help either.
...
-------------------------------------------------------------------------
iter correlation delta rms T2 Non-T2 Main
energy energy error ampl ampl Block
time time time
-------------------------------------------------------------------------
1 -1.1314730748 -1.131D+00 5.507D-02 19.33 0.00 12.95
2 -1.1704806544 -3.901D-02 2.914D-01 18.98 0.01 13.06
3 -1.1694014932 1.079D-03 8.983D-03 19.15 0.01 13.22
4 -1.1736828196 -4.281D-03 3.105D-03 19.07 0.01 13.08
5 -1.1750358517 -1.353D-03 2.367D-03 19.22 0.01 13.22
6 -1.1750616952 -2.584D-05 4.917D-03 19.15 0.01 13.15
...
20 -1.1750283271 9.944D-08 2.207D-03 19.39 0.01 13.38
21 -1.1750291661 -8.389D-07 9.038D-04 19.61 0.01 13.52
22 -1.1750320100 -2.844D-06 9.332D-04 19.29 0.01 13.28
23 -1.1750339244 -1.914D-06 9.920D-04 19.15 0.01 13.17
24 -1.1750237095 1.021D-05 1.520D-03 19.34 0.01 13.33
25 -1.1750229066 8.029D-07 2.125D-04 19.35 0.01 13.40
26 -1.1750153962 7.510D-06 7.016D-04 19.15 0.01 13.14
27 -1.1750172859 -1.890D-06 2.798D-03 19.09 0.01 13.16
...
I also tried to employ the TCE instead of the classical CC code of NWChem:
...
tce
freeze atomic
thresh 1.0e-06
maxiter 200
diis 5
lshift 0.0
2eorb
2emet 15
ccsd(t)
end
task tce energy
Same kind of convergence problem here, the CCSD residuum never reaches the default convergence threshold (1e-6), there seems to be a hard limit somewhere around 3e-5, also with different values for the "diis" and "lshift" parameters. The strange thing here is also that the residuum increases during the microcycles:
...
15 0.0058721120400 -1.1749892104618 41.5 41.7 10.1
MICROCYCLE DIIS UPDATE: 15 5
16 0.0015301379633 -1.1749241746419 42.9 43.2 10.3
17 0.0019761456618 -1.1749602667130 42.9 43.1 10.4
18 0.0031609178508 -1.1749794233376 43.1 43.3 10.2
19 0.0056086793834 -1.1749960594225 43.3 43.6 10.1
20 0.0105295930888 -1.1750045587169 42.8 43.0 10.4
MICROCYCLE DIIS UPDATE: 20 5
21 0.0009727911900 -1.1749396410455 43.1 43.3 10.2
...
55 0.0004747240041 -1.1750166044034 42.7 42.9 10.3
MICROCYCLE DIIS UPDATE: 55 5
56 0.0001816606161 -1.1750106882977 42.7 42.9 10.3
57 0.0001925416601 -1.1750145241198 43.0 43.2 10.2
58 0.0002210206154 -1.1750151612139 42.7 43.0 10.2
59 0.0002765153736 -1.1750162777920 42.8 43.1 9.9
60 0.0003889740905 -1.1750168095973 43.4 43.7 9.9
MICROCYCLE DIIS UPDATE: 60 5
61 0.0001419390062 -1.1750221743333 43.1 43.3 10.3
...
195 0.0000364980173 -1.1750178652350 42.3 42.6 10.3
MICROCYCLE DIIS UPDATE: 195 5
196 0.0000329190537 -1.1750178325642 42.1 42.3 10.5
197 0.0000329329384 -1.1750178699136 41.9 42.2 10.4
198 0.0000329950449 -1.1750178580599 42.0 42.3 10.5
199 0.0000331467713 -1.1750178626538 40.4 40.7 9.7
200 0.0000335334615 -1.1750178620145 40.5 40.7 9.6
I know that the problem I am having here in general disappears when the "set lindep:tol" parameter in line 2 of the input would be left at its default. But I need to keep this line to ensure that the HF, MP2, and CCSD(T) energies really correspond to the full aug-cc-pVDZ basis set, i.e., no elimination of near linear dependencies. The TCE even issues a warning when linear dependencies were removed "!! WARNING !! The number of MO is less than the number of AO".
Somebody with an idea how to get reasonable CCSD convergence behaviour for systems of the kind described here?
Sincerely,
Christian
|