Energy=0 for large df-ccsd(t) calculations

1

Hi,

Large df-ccsd(t) and fno-df-ccsd(t) calculations result in energy being zero.
see below for the output. Smaller df-ccsd(t) calculations are fine.
local gnu compilation with mkl.
Any ideas?

Let me know if you need additional info.

cheers,
Holger

                          Psi4 0.3.250 Driver

                      Git: Rev {master} 0081a7b dirty
[....]
  => Primary Basis Set <= 

  Basis Set: DEF2-SVP
Number of shells: 290
Number of basis function: 632
Number of Cartesian functions: 669
Spherical Harmonics?: true
Max angular momentum: 2

=> Auxiliary Basis Set <= 

Basis Set: DEF2-SVP-RI
Number of shells: 728
Number of basis function: 2098
Number of Cartesian functions: 2383
Spherical Harmonics?: true
Max angular momentum: 3

Number of auxiliary functions:        2098

==> Frozen Natural Orbitals <==

    Doubles contribution to MP2 energy in full space:      -5.693226392180

    Cutoff for significant NO occupancy: 1.000e-06

    Number of virtual orbitals in original space:    489
    Number of virtual orbitals in truncated space:   489

 ==> Memory <==

    Total memory available:          120162.96 mb

    CCSD memory requirements:         83064.81 mb
        3-index integrals:             6912.28 mb
        CCSD intermediates:           76152.53 mb

    (T) part (regular algorithm):     63331.10 mb

 ==> Input parameters <==

    Freeze core orbitals?                 yes
    Use frozen natural orbitals?          yes
    r_convergence:                  1.000e-07
    e_convergence:                  1.000e-06
    Number of DIIS vectors:                 4
    Number of frozen core orbitals:        41
    Number of active occupied orbitals:   102
    Number of active virtual orbitals:    489
    Number of frozen virtual orbitals:      0


 Begin singles and doubles coupled cluster iterations

 Iter  DIIS          Energy       d(Energy)          |d(T)|     time
  0   0 1    0.0000000000    0.0000000000    1.2897468632      440
  1   1 1    0.0000000000    0.0000000000    1.2897468632     9369
2

Can you provide the input.dat file?

3

yes sure, see below (not yet allowed to upload).

molecule upu {
-1 1
H    -5.0209632    2.6262239    0.9455459 
O    -4.2400389    2.1925254    0.5665500 
C    -4.4928824    1.9684076   -0.8288992 
H    -4.9542132    2.8514142   -1.2867260 
H    -5.1623462    1.1089123   -0.9716041 
C    -3.1941215    1.6639589   -1.5369371 
H    -3.4275656    1.4108960   -2.5787065 
O    -2.3219984    2.8346815   -1.5363557 
C    -0.9732214    2.4395291   -1.3824410 
H    -0.3479935    2.9173130   -2.1383561 
N    -0.4725116    2.9128802   -0.0516810 
C    -1.3240584    3.3308610    0.9391960 
H    -2.3784670    3.2537985    0.6965826 
C    -0.8802020    3.7962673    2.1338390 
H    -1.5774031    4.1214656    2.8944680 
C     0.5309825    3.8809477    2.3986800 
O     1.0703535    4.3034129    3.4270965 
N     1.3268504    3.4075228    1.3390642 
H     2.3345472    3.4471989    1.4684354 
C     0.9063554    2.9179838    0.1196198 
O     1.6837323    2.5410195   -0.7573277 
C    -2.3250237    0.5553668   -0.9379112 
H    -2.3312693    0.5959558    0.1558435 
C    -0.9420641    0.9129399   -1.4934760 
H    -0.1068781    0.4682769   -0.9474740 
O    -0.8623503    0.5879058   -2.8836835 
H    -1.4471213   -0.1810944   -3.0214125 
O    -2.7994431   -0.7224808   -1.3747659 
P    -2.6128469   -2.0172271   -0.3540581 
O    -3.3789706   -3.1203485   -1.0157780 
O    -2.8476897   -1.5817771    1.0575374 
O    -1.0014193   -2.3194568   -0.4384723 
C    -0.4478707   -2.7956984   -1.6817968 
H    -0.7137362   -2.1184432   -2.4994052 
H    -0.8437097   -3.7912688   -1.9094413 
C     1.0551746   -2.8794892   -1.5615131 
H     1.4468824   -3.2799378   -2.5061073 
O     1.6119912   -1.5540709   -1.3460484 
C     2.7626282   -1.6327928   -0.5224282 
H     3.6180955   -1.1609286   -1.0062693 
N     2.5124212   -0.8705839    0.7347134 
C     1.2474473   -0.7057556    1.2396101 
H     0.4554603   -1.1444551    0.6412561 
C     1.0042152   -0.0422655    2.3972099 
H    -0.0057338    0.0719565    2.7680647 
C     2.0984780    0.5049884    3.1598408 
O     2.0222681    1.1178911    4.2289523 
N     3.3565482    0.2799105    2.5732877 
H     4.1667216    0.6340971    3.0758497 
C     3.6353974   -0.3838691    1.3976802 
O     4.7795732   -0.5562671    0.9784469 
C     1.6046482   -3.7144327   -0.4042330 
H     1.0067330   -3.5684518    0.5007106 
C     3.0142712   -3.1185650   -0.2314534 
H     3.4356453   -3.2746626    0.7675919 
O     3.8813091   -3.6355402   -1.2401394 
H     3.5764918   -4.5520729   -1.4000962 
O     1.6266676   -5.0902294   -0.7840713 
H     1.6622231   -5.6301852    0.0233963 
}

set globals {
basis def2-SVP
#basis def2-TZVP

# DF
scf_type df
DF_BASIS_SCF def2-SVP-jkfit
df_basis_cc def2-svp-ri
guess read
#CD
#    scf_type cd
#    df_basis_cc cholesky

#DF_SCF_GUESS false
DIIS_MAX_VECS 4
freeze_core true
}

memory 126 Gb
set_num_threads(8)
energy('df-ccsd(t)')
4

I can try to take a look. Does it really take that much memory? Might be hard to test if so.

5

The example runs with just below 64Gb allocated virt mem, if one specifies 68Gb memory in the input (T2 on disk).
Below 68Gb psi4 complains and exits.

I just ran the example with the def2-SV§ basis set and it works fine, i.e. i get a non-zero energy.
for def2-SV§ the memory requirements are:

CCSD memory requirements:         62737.11 mb
    3-index integrals:             5749.08 mb
    CCSD intermediates:           56988.03 mb

maybe an integer overflow somewhere?

If anyone can give me tips where to start looking, I can try my luck as well, but i am more a fortran guy and some c++ things are confusing.

6

I can reproduce this on a 64gb machine. I’m sure it is an integer overflow, and I’m looking into it.

7

OK, I think I’ve fixed the problem. I haven’t run the job to convergence, but the MP2 at least agrees with that from df-mp2. I checked the code into psi4public, so grab a fresh copy and recompile. Please let me know if this problem persists.

8

hi, it seems indeed fixed! Thanks a lot. I get a normal looking CCSD correlation energy. The (T) is still running, in case it results to 0 or similar I will report it.

otherwise the topic can be considered closed.