cloning this to a separate thread so as not to clutter previous discussion.
Having to run larger, more expensive constrained DFT jobs with hybrid functionals so I started to explore the resolution of identity-based approximations. However, I find with the following input (RI_J and occ-RI-K) that it segfaults immediately before the first SCF cycle. Using Qchem 6.3.
$molecule
1 2
C 6.30855022 -4.93788338 3.88865279
H 6.46844030 -5.83477413 3.28091071
C 5.01765030 -4.35035744 4.19205135
H 4.04802354 -4.72207231 3.84570655
C 5.23841495 -3.22386205 5.02651884
H 4.47447698 -2.55909283 5.44169462
C 6.66721159 -3.10150946 5.22621529
H 7.15976745 -2.33556271 5.83473897
C 7.31655349 -4.20719489 4.58681439
H 8.38968765 -4.41966656 4.59195632
C 6.21501810 -1.01803699 2.56079378
H 6.09904537 -0.12748159 3.18572776
C 7.47643922 -1.62182529 2.21664122
H 8.46168129 -1.27259229 2.54082808
C 7.20721274 -2.73851668 1.37478405
H 7.94488276 -3.41511989 0.93265470
C 5.78075005 -2.82732645 1.20681162
H 5.27058208 -3.57749332 0.59529050
C 5.15516226 -1.69391917 1.85096660
C 3.74477713 -1.34129407 1.83063937
C 2.79744585 -2.10924529 1.10156345
H 3.10807348 -3.02219605 0.58703002
C 1.47285868 -1.69375406 0.97701026
H 0.77759409 -2.27246231 0.35960808
C 1.05307597 -0.50071272 1.58786675
C 1.94333084 0.24494950 2.36842866
H 1.63232837 1.17047784 2.85630207
C 3.26858932 -0.17875360 2.48969492
H 3.95611436 0.43480123 3.07808950
Co 6.25310142 -2.97668890 3.20704943
N -0.32023883 -0.01937991 1.29546237
C -0.56952759 1.42091634 1.63016136
H 0.20840818 2.02914160 1.15160308
H -1.55666329 1.68346606 1.22473574
H -0.56324389 1.55048325 2.72087926
C -1.37583337 -0.89912011 1.91359579
H -1.21877777 -1.93280239 1.57691648
H -1.28362197 -0.83601280 3.00726190
H -2.36068006 -0.53706103 1.58448065
H -0.41746428 -0.06422855 0.25424860
C 2.74081181 0.23358296 -1.33464848
C 3.07781445 1.33939183 -0.52471591
C 3.77785252 -0.57596110 -1.85603517
H 2.27171878 1.96727642 -0.13667488
H 3.54385475 -1.43198863 -2.49568935
C 4.41385497 1.62501708 -0.23428591
C 5.11313185 -0.29445065 -1.55455647
H 4.66306364 2.49129425 0.38751464
H 5.90839827 -0.93004235 -1.95696942
C 5.43472279 0.80285844 -0.73853255
H 6.48090493 1.02669988 -0.50553360
C 1.31075073 -0.06995661 -1.57864311
O 0.41861174 0.57454350 -0.99120839
C 0.93672357 -1.17041692 -2.54385239
H 1.37713004 -2.13610755 -2.23949946
H -0.15882613 -1.26433396 -2.58213330
H 1.32148929 -0.94112848 -3.55429865
$end
$rem
METHOD TPSSh
MEM_TOTAL 400000
JOBTYPE opt
BASIS def2-tzvpp
AUX_BASIS RIJK-def2-tzvpp
SCF_CONVERGENCE 8
XC_GRID 2 SG-2
RI_J TRUE
OCC_RI_K TRUE
RI_K_GRAD TRUE
MAX_SCF_CYCLES 200
GEOM_OPT_MAX_CYCLES 100
SOLVENT_METHOD smd
DFT_D D4
CDFT true
CDFT_THRESH 8
NO_REORIENT true
POINT_GROUP_SYMMETRY false
INTEGRAL_SYMMETRY false
MOLDEN_FORMAT TRUE
$end
$smx
solvent THF
$end
$opt
CONSTRAINT
stre 31 53 2.47538
ENDCONSTRAINT
$end
$cdft
-1.0
1 1 40
1.0
1 1 40 s
$end
/global/homes/s/foo/qchem: line 122: 2090781 Segmentation fault ${QCPROG_S} ${inp} ${scr}
Error in Q-Chem run part 1
When the memory is lowered to 4K (in case it’s allocated per thread), a different segfault results:
corrupted double-linked list
/global/homes/s/foo/qchem: line 122: 2357577 Aborted ${QCPROG_S} ${inp} ${scr}
Error in Q-Chem run part 1
Error in the Q-Chem run
Last time I tried this with def2-SVP on 6.2.1, it barely squeaks out the first SCF cycle before a similar segfault.
With RI_J set to false (so only occ-RI-K and the corresponding gradient) the job runs but very slowly, 8 hours for 2 converged SCF cycles. With def2-SVP, it’s about an hour per converged SCF cycle. The analytical gradients didn’t seem to be available in any of these cases for CDFT.
Just running the hybrid functional CDFT without the RI approximations, the opt job runs and completes normally with def2-TZVPP and takes about 20 node-hours for 25 geometry steps. I could afford to use it for a few single point calculations but not for the full dataset.
Possibly all of this is a CDFT issue so I am trying a test job without.