Different CCSD correlation energies between TCE and legacy CCSD code


Clicked A Few Times
I noted that a similar question has been asked by others before, but there is no good answer found yet:

I have been doing a comparison on the performances and resource demands of CCSD various software package; I found that NWChem may give different CCSD energies for using the TEC and legacy CCSD algorithm, and I reproduced the number generated by the legacy CCSD code by other commercial packages. Can anyone provide some advice?


The code was download and compiled on 16May2019, using intel compiler, mpi and mkl (2017), the armci_network used is mpi-pr, and ga version is 5.7.0


The summary of outputs and input are given as follow:

Legacy CCSD:
-------------------------------------------------------------------------
iter correlation delta rms T2 Non-T2 Main
energy energy error ampl ampl Block
time time time
-------------------------------------------------------------------------
g_st2 size: 12 MB
mem. avail 10099 MB
1 -0.8027540481 -8.028D-01 1.771D-01 4.70 0.00 4.36
...
11 -0.8316067315 -2.419D-07 4.788D-07 1.72 0.00 1.46
*************converged*************
-------------------------------------------------------------------------

-----------
CCSD Energy
-----------
Reference energy: -230.711542913190101
CCSD corr. energy: -0.831606731454189
Total CCSD energy: -231.543149644644302


TCE CCSD:
 CCSD iterations
-----------------------------------------------------------------
Iter Residuum Correlation Cpu Wall V2*C2
-----------------------------------------------------------------
1 0.2204269357423 -0.8055763310885 5.2 1.4 0.8
...
35 0.0000000000008 -0.8473895042521 4.0 1.0 0.7
-----------------------------------------------------------------
Iterations converged
CCSD correlation energy / hartree = -0.847389504252062
CCSD total energy / hartree = -231.558932417442151


The input for NWChem:
memory stack 10000 mb heap 100 mb global 10000 mb noverify
start benzene
geometry
SYMMETRY c1
C 0.00000 1.40272 0.00000
H 0.00000 2.49029 0.00000
C -1.21479 0.70136 0.00000
H -2.15666 1.24515 0.00000
C -1.21479 -0.70136 0.00000
H -2.15666 -1.24515 0.00000
C 0.00000 -1.40272 0.00000
H 0.00000 -2.49029 0.00000
C 1.21479 -0.70136 0.00000
H 2.15666 -1.24515 0.00000
C 1.21479 0.70136 0.00000
H 2.15666 1.24515 0.00000
end

basis cartesian
* library 6-31g**
end

driver
clear
end

scf
singlet
VECTORS HCORE
THRESH 1.0e-9
maxiter 90
end

ccsd
maxiter 100
TOL2E 10e-12
FREEZE atomic
end

TCE
FREEZE atomic
ccsd(t)
thresh 1.0e-12
io ga
tilesize 18
END

task ccsd(t) energy
task tce energy
set tolguess 1e-7
set tce:xmem 100

Regards,
Dominic Chien

Forum Vet
Your TCE CCSD result is quite different, thus perhaps it is wrong. Please note the threshes for both are different when you get the results. I cannot understand why you use tolguess 1e-7 to make the input a little confusing. The two CCSDes did not converge at the same thresh.
I can say at least for this molecule the two CCSD and CCSD(t) energies are almost identical when using NWCHEM6.8 because I have tried.

Forum Vet
I have put the log files on Github.


Forum >> NWChem's corner >> Running NWChem