Weird Energy with DF-CCSD(T) on pre-built 1.2 binary


#1

Dear all,

I’m attempted a DF-FNO-CCSD(T)/aug-cc-pvdz computation on sumanene. I’m getting an asinine energy. If I use cc-pvdz I get a normal energy that is about the same as another program’s CCSD energies. I installed stable release 1.2.1 and the python 3.5 version of it for linux. The hardware is intel haswell cpu’s.

 0   0 1 82301976411.2597808838 82301976411.2597808838 180071.6863671333       79
  1   1 1 -11765764147947413438464.0000000000 -11765764148029716168704.0000000000 5189755552069629.0000000000     1039
  2   2 1 18577504081805034984135645462528.0000000000 18577504093570799348831936839680.0000000000 13953090784875981570048.0000000000     1042
  3   3 1 -10975754574997206597632.0000000000 -18577504092780789161197672857600.0000000000 2850375421929338594918400.0000000000     1058

Here’s the input file

memory 120 GB
molecule {
0 1
C 3.269671140 0.262382661 0.000000000
C -1.638988443 0.201021220 2.834237354
C -1.638988443 0.201021220 -2.834237354
H 3.645686428 1.289803000 0.000000000
H -1.846170088 1.221151217 3.170972567
C 1.219967165 -0.666672220 0.713456211
C 0.020624067 -0.681664789 1.405951372
C 0.020624067 -0.681664789 -1.405951372
C 1.219967165 -0.666672220 -0.713456211
C -1.215021798 -0.697111167 -0.692495161
C -1.215021798 -0.697111167 0.692495161
C 2.350712693 -0.010518009 1.202305693
C -0.133462307 -0.041571814 2.636656855
C -0.133462307 -0.041571814 -2.636656855
C 2.350712693 -0.010518009 -1.202305693
C -2.215754163 -0.067601818 -1.434351162
C -2.215754163 -0.067601818 1.434351162
C 2.253218542 0.518128480 2.489250128
C 1.019790334 0.502709825 3.201425874
C 1.019790334 0.502709825 -3.201425874
C 2.253218542 0.518128480 -2.489250128
C -3.291380529 0.448817382 -0.712175746
C -3.291380529 0.448817382 0.712175746
H 3.077229877 1.077986926 2.924280546
H 0.974199492 1.051697677 4.138560570
H 0.974199492 1.051697677 -4.138560570
H 3.077229877 1.077986926 -2.924280546
H -4.090407292 0.988386795 -1.214280023
H -4.090407292 0.988386795 1.214280023
H 4.154760170 -0.387272435 0.000000000
H -2.069145762 -0.465075312 3.593654516
H -2.069145762 -0.465075312 -3.593654516
H -1.846170088 1.221151217 -3.170972567
}

set {
basis aug-cc-pvdz
guess sad
reference rhf
scf_type DF
cc_type DF
freeze_core true
OCC_TOLERANCE 1.0e-5
}

energy(‘fno-ccsd(t)’)


#2

I get the same with the latest developer version. Furthermore it is not caused by the FNO approx, the FNO part looks normal and DF-CCSD(T) gives the same nonsense.

The DF tensor could have some linear dependencies perhaps? The conventional FNO-CCSD(T) seems to work:

   Iter  DIIS          Energy       d(Energy)          |d(T)|     time
      0   0 1   -2.8444805434   -2.8444805434    0.0447735979      330
      1   1 1   -2.8491945311   -0.0047139877    0.0190091062      344
      2   2 1   -2.8540707763   -0.0048762452    0.0088988976      323

#3

+1 for linear depends in the inverse coulomb metric. Best to raise this on GitHub so the developers can add this to a TODO list.


#4

@bkw Do you have a github account and want to write an issue? If not I can make a post for you. Please include link to this forum post.


#5

@hokru I’ll get on it.


#6

@hokru I posted this to github


#7

What if you set a smaller value for DF_FITTING_CONDITION? Try 1e-7.


#8

Oh but you’ll have to do this on 1.3 since 1.2 isn’t passing the value properly.


#9

Would you recommend I grab the nightly version then?


#10

Not yet. There is a patch posted that might fix the problem, but it is not included yet.

Meanwhile you can run with conventional integrals cc_type conv. The integrals transformation takes some additional time, but overall it wont be much slower for this problem size.


#11

Well, it looks like somehow symmetry is the problem.

You can run your current calculation setting symmetry c1 in the geometry block. As DF-FNO-CCSD(T) does not use symmetry anyway, this is essentially no loss in speed.

Please try and let us know.


#12

@hokru I’m running it now

@hokru disabling symmetry for the scf seems to fix the issue. The energies are about the same as MRCC’s DF-CCSD(T) code. Thank you for the assistance.