Convergence issues

Forum Vet

9:18:15 AM PDT - Tue, Mar 26th 2013

Dear NWChem developers and community members,

I’m running into a problem that I am unable to resolve and would like to ask for your help. The problem is that I’m generally unable to achieve SCF convergence on relatively simple organic molecules when using a basis with diffuse functions. Specifically, basis sets of the aug-pc-X and aug-cc-pVXZ type are problematic. I’m encountering the problem with NWChem versions 6.0, 6.1, and 6.1.1.

To help with convergence, I first run an energy calculation with the basis set without diffuse functions, then read in the converged vectors for a second energy calculation with diffuse functions added. Unfortunately, this has not helped in the majority of cases. The calculation without diffuse functions runs fine, but the calculation with the diffuse functions has a large DIIS error from the start on and typically starts to diverge after a few iterations (below is an example).

d= 0,ls=0.0,diis     1   -539.3212395974 -1.31D+03  7.60D-01  9.07D+00   148.1

d= 0,ls=0.0,diis     2   -539.3492153346 -2.80D-02  9.04D-04  7.74D+03   203.8

d= 0,ls=0.0,diis     3   -539.3492043060  1.10D-05  5.63D-02  7.74D+03   259.4

d= 0,ls=0.0,diis     4   -539.3468186446  2.39D-03  3.37D-03  7.75D+03   315.1

d= 0,ls=0.0,diis     5   -539.3474219125 -6.03D-04  2.94D-02  7.75D+03   370.9

d= 0,ls=0.0,diis     6   -539.3484646008 -1.04D-03  5.10D-02  7.73D+03   426.6

d= 0,ls=0.0,diis     7   -539.3490772651 -6.13D-04  4.11D-02  7.73D+03   482.3

d= 0,ls=0.0,diis     8   -539.3475847257  1.49D-03  8.91D-02  7.73D+03   538.7

d= 0,ls=0.0,diis     9   -539.3447532079  2.83D-03  7.63D-03  7.73D+03   594.5

d= 0,ls=0.0,diis    10   -539.3450053401 -2.52D-04  7.80D-02  7.73D+03   650.2

d= 0,ls=0.0,diis    11   -539.3453092257 -3.04D-04  1.02D+04  7.74D+03   705.8

Grid integrated density:      94.000042075022

Requested integration accuracy:   0.10E-07

d= 0,ls=0.5,diis    12  -3005.8035648386 -2.47D+03  8.44D+02  3.75D+11   764.3

Grid integrated density:      94.000041943620

Requested integration accuracy:   0.10E-07

d= 0,ls=0.5,diis    13   -337.1543884878  2.67D+03  1.24D+03  7.09D+11   822.8

Grid integrated density:      94.000035846688

Requested integration accuracy:   0.10E-07

d= 0,ls=0.5,diis    14    375.2187617181  7.12D+02  1.09D+04  7.66D+10   881.2

Grid integrated density:      94.000021289358

Requested integration accuracy:   0.10E-07

d= 0,ls=0.5,diis    15    460.6161514873  8.54D+01  4.54D+03  4.84D+10   939.7

Grid integrated density:      94.000003943278

Requested integration accuracy:   0.10E-07

d= 0,ls=0.5,diis    16   -518.7285508120 -9.79D+02  3.60D+02  9.40D+06   998.2

Grid integrated density:      94.000003158007

Requested integration accuracy:   0.10E-07

d= 0,ls=0.5,diis    17   -533.3548415213 -1.46D+01  1.36D+02  1.18D+07  1056.7

Grid integrated density:      94.000003137809

Requested integration accuracy:   0.10E-07

d= 0,ls=0.5,diis    18   -430.1794423581  1.03D+02  5.55D+02  5.74D+09  1115.2

Grid integrated density:      94.000003397106

Requested integration accuracy:   0.10E-07

d= 0,ls=0.5,diis    19   -542.7589569074 -1.13D+02  3.70D+03  1.91D+07  1173.7

Grid integrated density:      94.000001434088

Requested integration accuracy:   0.10E-07

d= 0,ls=0.5,diis    20   -532.1381465702  1.06D+01  8.53D+00  2.94D+04  1231.3

Grid integrated density:      94.000001399588

Requested integration accuracy:   0.10E-07

d= 0,ls=0.5,diis    21   -531.7068488856  4.31D-01  1.79D+01  2.35D+04  1289.0

d= 0,ls=0.5,diis    22   -532.9625586983 -1.26D+00  4.52D+03  7.90D+03  1345.4

I’ve tried many things in order to improve convergence, such as increasing the accuracy of various cutoffs, improving the DFT grid, level shifting, dampening, all to no avail.

Strangely enough, however, I found that for many cases specifying “set lindep:n_dep 0” will cause the calculation to converge without a problem (provided the degree of partial linear dependence isn’t too large). When setting “set lindep:n_dep” or “set lindep:tol” such that only one vector is removed prior to the second calculation, the problem immediately reappears.

To me this makes no sense, as I understood that SCF convergence should usually improve upon removing linear dependent vectors, not worsen. Did I perhaps encounter a bug? Or did I overlook an important setting?

I would greatly appreciate if someone could help me with getting the calculation to convergence without having to set “set lindep:n_dep 0”. Below is the input of my calculation. Thank you in advance.

Kind regards,
Thom

start phenanthrene

geometry

 C                  0.00000000    2.84859800    0.88680600

 C                  0.00000000    3.57955900   -0.29438700

 C                  0.00000000    2.89936800   -1.53469700

 C                  0.00000000    1.50842000   -1.57667000

 C                  0.00000000    0.73043500   -0.38774400

 C                  0.00000000    1.42795200    0.87014000

 C                  0.00000000   -0.73043500   -0.38774400

 C                  0.00000000   -1.42795200    0.87014000

 C                  0.00000000   -0.68487000    2.10145900

 C                  0.00000000    0.68487000    2.10145900

 C                  0.00000000   -1.50842000   -1.57667000

 C                  0.00000000   -2.89936800   -1.53469700

 C                  0.00000000   -3.57955900   -0.29438700

 C                  0.00000000   -2.84859800    0.88680600

 H                  0.00000000    3.35683000    1.85437700

 H                  0.00000000    4.67108900   -0.26819800

 H                  0.00000000   -1.24321500    3.04075300

 H                  0.00000000    1.24321500    3.04075300

 H                  0.00000000   -1.01691100   -2.54926800

 H                  0.00000000   -3.46767400   -2.46710500

 H                  0.00000000   -4.67108900   -0.26819800

 H                  0.00000000   -3.35683000    1.85437700

 H                  0.00000000    3.46767400   -2.46710500

 H                  0.00000000    1.01691100   -2.54926800

 symmetry c2v

end

basis "small" spherical

 H library cc-pVDZ

 C library cc-pVDZ

end

basis "large" spherical

 H library aug-cc-pVDZ

 C library aug-cc-pVDZ

end

set "ao basis" small

scf

 tol2e 1e-12

end

dft

 xc becke88 1.0 lyp 1.0

 grid xfine

 tolerances tight

 direct

 iterations 100

 vectors input atomic output smallbasis.movecs

end

task dft energy

set "ao basis" large

dft

 xc becke88 1.0 lyp 1.0

 grid xfine

 tolerances tight

 direct

 iterations 100

 vectors input project "small" smallbasis.movecs output largebasis.movecs

end

task dft energy

Just Got Here

9:26:03 AM PDT - Tue, Mar 26th 2013
Dear all, I should clarify that the above message was written by me. However, as I was (for some unknown reason) unable to post it, Bert was kind enough to do so for me. Kind regards, Thom

Forum Vet

9:56:37 AM PDT - Tue, Mar 26th 2013

It seems that the projection from a small basis to a large basis with linear dependencies is creating some problems. I don't know why you have to do the projection because either the cc-pVDZ or aug-cc-pVDZ converge in 9 iterations starting from a default atomic guess, using the input below.

Bert

start phenanthrene
geometry

 C                  0.00000000    2.84859800    0.88680600

 C                  0.00000000    3.57955900   -0.29438700

 C                  0.00000000    2.89936800   -1.53469700

 C                  0.00000000    1.50842000   -1.57667000

 C                  0.00000000    0.73043500   -0.38774400

 C                  0.00000000    1.42795200    0.87014000

 C                  0.00000000   -0.73043500   -0.38774400

 C                  0.00000000   -1.42795200    0.87014000

 C                  0.00000000   -0.68487000    2.10145900

 C                  0.00000000    0.68487000    2.10145900

 C                  0.00000000   -1.50842000   -1.57667000

 C                  0.00000000   -2.89936800   -1.53469700

 C                  0.00000000   -3.57955900   -0.29438700

 C                  0.00000000   -2.84859800    0.88680600

 H                  0.00000000    3.35683000    1.85437700

 H                  0.00000000    4.67108900   -0.26819800

 H                  0.00000000   -1.24321500    3.04075300

 H                  0.00000000    1.24321500    3.04075300

 H                  0.00000000   -1.01691100   -2.54926800

 H                  0.00000000   -3.46767400   -2.46710500

 H                  0.00000000   -4.67108900   -0.26819800

 H                  0.00000000   -3.35683000    1.85437700

 H                  0.00000000    3.46767400   -2.46710500

 H                  0.00000000    1.01691100   -2.54926800

 symmetry c2v

end

basis "small" spherical

 H library cc-pVDZ

 C library cc-pVDZ

end

basis "large" spherical

 H library aug-cc-pVDZ

 C library aug-cc-pVDZ

end

set "ao basis" small

dft

 xc becke88 1.0 lyp 1.0

 tolerances tight

 direct

 iterations 100

end

task dft energy

set "ao basis" large

dft

 xc becke88 1.0 lyp 1.0

 tolerances tight

 direct

 iterations 100

end

task dft energy

Just Got Here

10:11:37 AM PDT - Tue, Mar 26th 2013
Dear Bert, Thank you for your reply. The reason I do the projection is because of SCF convergence problems with some of the larger molecules that I would like to calculate on. At least, starting directly from an atomic guess with a larger augmented basis such as aug-cc-pVTZ didn't work for those molecules, which I why I chose to do projection. I agree that projection is not needed for the above example (which was provided solely for demonstrating the projection problem). Do you know if I can resolve this problem with the projection in some way? Kind regards, Thom

Forum Vet

10:42:48 AM PDT - Tue, Mar 26th 2013

Can you provide an input example of a real system you want to do that fails to converge? This case above is fine with a aug-cc-pVTZ basis set.

One trick I can suggest is to run a Hartree-Fock calculation first and use those vectors as a start for the DFT. An example:

scf

 vectors output test.movecs

end

task scf

dft

 vectors input test.movecs output dft.movecs

end

task dft

Just Got Here

2:35:58 AM PDT - Thu, Mar 28th 2013
I've tried using a converged Hartree-Fock calculation as starting vectors with a few of my molecules and found that this works well. Thank you for this suggestion. Kind regards, Thom

Just Got Here

6:15:27 AM PDT - Mon, Jun 29th 2020
SCF convergence Failure
Hi , I have been trying to find the ESP charges of NbO3SiO2- molecule but it fails to converge all the time. Can someone help me with this problem? Here is my input file. kind regards, Sahan title cluster1_dft_opt echo print medium geometry noautosym noautoz Nb 0.000000 0.000000 0.000000 O 0.000000 0.000000 1.820000 O 1.576166 0.000000 -0.910000 O -1.818653 0.000000 -1.050000 Si -3.179196 0.000000 0.113196 O -4.725403 0.000000 -0.433471 O -3.355789 -1.342500 1.283926 end basis Nb library lanl2dz_ecp O library 6-31+G* Si library 6-31+G* end ecp Nb library lanl2dz_ecp end charge -1 dft iterations 200 xc b3lyp decomp end task dft optimize task esp

Forum Vet

12:02:16 PM PDT - Mon, Jun 29th 2020
Sahan92 Please post your NWChem questions to the Google Group, since this website will be shutdown very soon. Here is the link for the new forum https://groups.google.com/forum/#!forum/nwchem-forum

Forum >> NWChem's corner >> Running NWChem