# Keywords ignored!

3 years 11 months ago - 3 years 11 months ago #451 by Nike
Keywords ignored! was created by Nike
Greetings!
When the keywords are printed out, you can see that I've requested for a maximum of 999 cc iterations:
Keywords:

active=none
agrid=ld0006-ld0590
basis=acv6z-koput
basopt=off
bpcompo=0.985
bpcompv=0.98
bpdfo=0.985
bpocc=0.985
bppdo=0.985
bppdv=0.98
bpedo=0.985
bpedv=0.98
calc=ccsdt(q)
ccmaxit=999

But it stopped after 50, even though it wasn't converged !!
======================================================================
Norm of residual vector:   276.83385295
CPU time [min]:  7508.604                   Wall time [min]: 18841.375

Iteration 49  CC energy:   -14.95593221  Energy decrease:   0.00000050
======================================================================
Norm of residual vector:   269.39534350
CPU time [min]:  7657.753                   Wall time [min]: 19077.904

Iteration 50  CC energy:   -14.95595314  Energy decrease:   0.00002093
======================================================================

Convergence not achieved in 50 iterations!

I've been getting these huge residuals whenever I work with a big basis set.
For ccsdt(q) my strategy is to start by using the hand-coded ccsd program rather than the automated mrcc program (then use rest=1 with fort.17 to get ccsdt(q) from a good initial guess).
However I'm getting huge norms for ccsd as well, so my thought was to use an MP2 guess for ccsd, but I think that's already the default:

"If you are still interested in the MP2 energy without DF, you can, e.g., run a CCSD calculation (without DF), where the MP2 energy is also calculated."

But I don't see MP2 printed anywhere. Is it possible to get it printed when it's used as an initial guess for ccsd?

So in summary:
1) MRCC seems to be ignoring my ccmaxit keyword
2) Therefore I want to try to converge ccsd in fewer iterations, is it already using the best possible initial guess (for example MP2, even though it is not printed)?
3) If it's already using the best initial guess, is it possible to play around with the DIIS parameters / damping / level shifts / etc. for getting better ccsd convergence?
4) I note that CFOUR successfully converged CCSDT for this basis set, but cannot do CCSDT(Q) for open-shell systems, so is it possible for MRCC to read in CCSDT amplitudes calculated from CFOUR? I guess the CFOUR amplitdues would have to be converted into fort.17 format somehow.

With best wishes !!!
Nike
Last edit: 3 years 11 months ago by Nike.

3 years 11 months ago #454 by Nike
Greetings,
I will find it very difficult to justify publishing this paper without doing CCSDT(Q) because it only requires 361GB of RAM and we have at least 48 machines with 0.5TB, 24 machines with 1.5TB and ~10 machines with 3TB of RAM. Furthermore the difference between CCSDT and CCSDT(Q) in all the smaller basis sets is far bigger than our spectroscopic requirements, but the difference between CCSDT(Q) and FCI is two orders of magnitude smaller, and well within the estimated basis set error.

However CFOUR cannot do CCSDT(Q) for triplet states, and MRCC cannot seem to do even CCSD for this system of only 6 electrons.

CCSDT(Q) with rest=2 starts with CCSD, but it doesn't converge, and my keyword ccmaxit=999 was ignored (!) so only 50 iterations are done before the parenthesis T correction:

Norm of residual vector: 1.52164432
Norm of residual vector: 103.47784715
Norm of residual vector: 1111.73461563
Norm of residual vector: 1108.54662727
Norm of residual vector: 2716.46073811
Norm of residual vector: 2717.12325935
Norm of residual vector: 4609.78583500
Norm of residual vector: 4613.31762382
Norm of residual vector: 1955.46391930
Norm of residual vector: 1947.67262723
Norm of residual vector: 1385.92984900
Norm of residual vector: 1356.87944936
Norm of residual vector: 330.50116305
Norm of residual vector: 359.70279004
Norm of residual vector: 313.47016908
Norm of residual vector: 273.01953843
Norm of residual vector: 441.70573682
Norm of residual vector: 445.22843261
Norm of residual vector: 2373.48515483
Norm of residual vector: 2379.63700509
Norm of residual vector: 781.82795335
Norm of residual vector: 781.52418902
Norm of residual vector: 1138.33167826
Norm of residual vector: 1111.55570394
Norm of residual vector: 561.18349366
Norm of residual vector: 561.08001835
Norm of residual vector: 555.44852163
Norm of residual vector: 560.68389954
Norm of residual vector: 568.03771132
Norm of residual vector: 631.16346521
Norm of residual vector: 635.86403293
Norm of residual vector: 569.33756800
Norm of residual vector: 582.60058524
Norm of residual vector: 584.78689152
Norm of residual vector: 1392.17812921
Norm of residual vector: 1404.13194475
Norm of residual vector: 318.31944027
Norm of residual vector: 314.83076043
Norm of residual vector: 298.82961261
Norm of residual vector: 298.96959495
Norm of residual vector: 294.76510968
Norm of residual vector: 294.52263458
Norm of residual vector: 298.17524547
Norm of residual vector: 293.66865473
Norm of residual vector: 249.29143373
Norm of residual vector: 247.15765128
Norm of residual vector: 239.11450594
Norm of residual vector: 236.23230762
Norm of residual vector: 276.83385295
Norm of residual vector: 269.39534350

Then CCSDT is even worse:

Norm of residual vector: 216.05501167
Norm of residual vector: **************
Norm of residual vector: **************
Norm of residual vector: 69338.35229423
Norm of residual vector: 69440.70371954
Norm of residual vector: 83906.75572433
Norm of residual vector: 82617.09148452
Norm of residual vector: 21438.60809680
Norm of residual vector: 21210.75293370
Norm of residual vector: 19390.20275973
Norm of residual vector: 17406.63327475
Norm of residual vector: 4361.77005132
Norm of residual vector: 5007.35971420
Norm of residual vector: 6058.39726897
Norm of residual vector: 5981.96870782
Norm of residual vector: 4706.81912661
Norm of residual vector: 10276.10532990
Norm of residual vector: 11156.13943114
Norm of residual vector: 6861.12313291
Norm of residual vector: 6581.58471916
Norm of residual vector: 9197.54310257
Norm of residual vector: 9048.96189291
Norm of residual vector: 6807.21499107
Norm of residual vector: 13377.66515508
Norm of residual vector: 12860.53517424
Norm of residual vector: 19913.61125985
Norm of residual vector: 19585.77031727
Norm of residual vector: 13170.96989695
Norm of residual vector: 12384.07292072
Norm of residual vector: 4457.20613536
Norm of residual vector: 4543.36381345
Norm of residual vector: 3598.65359654
Norm of residual vector: 3688.20150008
Norm of residual vector: 4062.44116064
Norm of residual vector: 4218.69339575
Norm of residual vector: 4399.82094583
Norm of residual vector: 4667.99116302
Norm of residual vector: 4511.18536183
Norm of residual vector: 5156.31651095
Norm of residual vector: 5011.73598029
Norm of residual vector: 4854.11948010
Norm of residual vector: 5115.78221105
Norm of residual vector: 4133.87714528
Norm of residual vector: 5657.19131633
Norm of residual vector: 5654.22585608
Norm of residual vector: 5313.03206130
Norm of residual vector: 5274.61178229
Norm of residual vector: 6705.13562045
Norm of residual vector: 6762.62024000

Then I tried with the hand-coded CCSD program instead of the automated general MRCC program, but the norms were no smaller (in fact they were usually a bit bigger):

Norm of residual vector: 1.52164432
Norm of residual vector: 103.47784701
Norm of residual vector: 1111.73474752
Norm of residual vector: 1108.54675967
Norm of residual vector: 2716.46077697
Norm of residual vector: 2717.12319571

Therefore I wonder how I might be able to get CCSDT(Q) for this system.
For example:

1) getting more than 50 iterations in the CC?
2) starting with an MP2 initial guess (or is that already the case?) ?
3) Play around with the DIIS parameters, damping, level shifts, etc. ?
4) CCSDT converged easily in CFOUR and CCSD(T) worked in CFOUR and MOLRPO, can MRCC read the cluster amplitudes somehow, if I gt those in fort.17 format?
5) Can we project the cluster amplitudes from a converged calculation in a smaller basis, into the bigger basis, which is what we do at HF level when we use scfiguess=restart and the file SCFDENSITIES?

Any advice would be highly appreciated.

The MINP is as follows:

basis=aCV6Z-KOPUT
uncontract=off
iface=cfour
calc=CCSDT(Q)
mem=768GB
core=corr
cctol=5
ccmaxit=999
scfmaxit=9999
scftype=ROHF
scfiguess=ao
rohftype=semicanonical
rest=2
refdet=serialno
1,2
3,4

mult=3
symm=6
geom
Li
Li 1 R

R=4.1700

unit=angstroms

GENBAS:

LI:aCV6Z-KOPUT
EMSL BASIS SET LIBRARY

7
0 1 2 3 4 5 6
13 12 10 8 6 4 2
22 15 10 8 6 4 2

103207.000000 15664.800000 3600.840000 1026.350000 335.925000 121.409000 47.337200 19.630900 8.547060 3.853030 1.780350 0.837931 0.390527 0.097776 0.044933 0.020602 0.006100 49.390700 26.340100 14.047200 7.491400 3.995100

0.000004 0.000001 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.000029 0.000005 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.000149 0.000023 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.000627 0.000098 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.002274 0.000355 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.007358 0.001159 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.021450 0.003373 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.055599 0.008978 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.124328 0.020423 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.228994 0.040705 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.323811 0.064625 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.000000 0.000000 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.000000 0.000000 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.000000 0.000000 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.000000 0.000000 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.000000 0.000000 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0
0.000000 0.000000 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0
0.000000 0.000000 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0
0.000000 0.000000 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0
0.000000 0.000000 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0
0.000000 0.000000 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0
0.000000 0.000000 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0

32.644300 7.758170 2.461690 0.869615 0.327205 0.134815 0.059995 0.028101 0.013320 0.003900 51.457700 23.052200 10.327000 4.626300 2.072500

0.000229 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.001768 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.007560 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.023322 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.000000 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.000000 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.000000 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.000000 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.000000 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0
0.000000 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0
0.000000 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0
0.000000 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0
0.000000 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0
0.000000 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0
0.000000 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0

0.825300 0.437300 0.231700 0.122800 0.065000 0.034500 34.914700 15.004700 6.448300 2.771200

1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0

0.679100 0.358800 0.189500 0.100100 0.052900 23.047600 8.994200 3.509900

1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0

0.419900 0.230900 0.127000 0.069900 15.913400 5.499300

1.0 0.0 0.0 0.0 0.0 0.0
0.0 1.0 0.0 0.0 0.0 0.0
0.0 0.0 1.0 0.0 0.0 0.0
0.0 0.0 0.0 1.0 0.0 0.0
0.0 0.0 0.0 0.0 1.0 0.0
0.0 0.0 0.0 0.0 0.0 1.0

0.337100 0.183500 0.099900 10.203500

1.0 0.0 0.0 0.0
0.0 1.0 0.0 0.0
0.0 0.0 1.0 0.0
0.0 0.0 0.0 1.0

0.292700 0.146200

1.0 0.0
0.0 1.0

With best wishes!!
Nike

3 years 11 months ago #455 by kallay
Dear Nike,
1) The huge residual norms clearly indicate that something is wrong with the MO integrals you get from Cfour. You should check your Cfour input and output file if everything is correct. You should also try to run this calculation with the same settings but smaller basis sets.
Preferably, you should run this calculation with mrcc in standalone mode. I have run a couple of calculations with smaller basis sets and have not experience any problem.
You can change the max. number of iterations by changing the maxit variable in the MRCCCOMMON file and recompiling the code, but it will not help.
2) This is the case.
3) It is not possible but would not help.
4) It is not possible.
5) It is not possible

Best regards,
Mihaly Kallay

3 years 11 months ago #456 by Nike
Dear Mihaly,
Thank you VERY much for the reply.
- All integrals were done in MRCC.
- iface=cfour was just to trick MRCC into reading it's own fort.55.
- I have indeed converged CCSDTQ in aCV5Z with standalone MRCC, but CCSD in aCV6Z gives a huge norm.
- I have experienced huge residual norms with all other cases where I used a bit basis set in standalone MRCC (also for other molecules).
- Considering that CFOUR converges CCSDT for this aCV6Z basis set, and CFOUR and MRCC both have the same default itol=13, I suspect there is a bug in MRCC.

However I'll try it again in standalone MRCC with itol=18 and more a strict convergence criterion for the ROHF and let you know if it helps.

Since you say "You should check your Cfour input and output file if everything is correct," I have pasted below, the MRCC input file that was used to generate the fort.55.

basis=aCV6Z-KOPUT
uncontract=off
calc=CCSDT(Q)
mem=490GB
core=corr
cctol=7
ccmaxit=999
scfmaxit=9999
scftype=ROHF
scfiguess=ao
rohftype=semicanonical
mult=3
rest=2
geom
Li
Li 1 R

R=4.1700

unit=angstroms

I have some short questions about the other points:
2) Since MP2 is used for the initial guess, is there a way to print it, so I can compare it with other codes?
3) Ok.
4) Ok.
5) Is reading in cluster amplitudes from a smaller basis set, and projecting them to a bigger basis set, something that's impossible mathematically, or is it just not implemented? I note that (3) and (4) are of course possible in theory, but just not implemented in MRCC. I wonder if the same is true for (5) or if there is *really* no method (derived yet) for doing the projection.

With best wishes!
Nike

3 years 11 months ago #457 by kallay
Dear Nike,
Did the SCF converge for your aCV6Z calculation? Could you post the entire output file for the aCV6Z calculation?
2) The MP2 energy is printed out for canonical HF references (RHF/UHF) but not calculated for ROHF.
5) It is possible but not implemented.

Best regards,
Mihaly Kallay

3 years 11 months ago #460 by Nike
Dear Mihaly,
Thanks very much for the reply!
I guess the reason why the CCSD is starting with a huge residual norm is because MP2 isn't calculated for ROHF so the initial guess is not so good.

The ROHF converged:
ITERATION STEP  22
CPU time [min]:    12.536                   Wall time [min]:    11.123

ALPHA OCC:   2   0   0   0   0   2   0   0
BETA  OCC:   1   0   0   0   0   1   0   0
***HARTREE-FOCK ENERGY IN STEP  22 IS     -14.8640195364116519 [AU]
======================================================================

SUCCESS...
THE SCF ITERATION HAS CONVERGED!

Ag  B1g B2g B3g Au  B1u B2u B3u
FINAL ALPHA OCC:   2   0   0   0   0   2   0   0
FINAL BETA  OCC:   1   0   0   0   0   1   0   0

***FINAL HARTREE-FOCK ENERGY:    -14.8640195364116519 [AU]
***SEMICANONICAL ROHF ENERGY:    -14.8640195364113339 [AU]

But this was with the default scftol and itol.
I have attached various output files.

out.1462030 is where the ROHF converged, but CC didn't start, probably due to lack of RAM
out.1482423 used the fort.55 from the previous run and 50 iterations of CCSD was not enough to converge. I increased the RAM from 490GB to 768GB here, which I was nervous about because sometimes this causes "The number of active orbitals has been changed! It is dangerous to restart the program!" but I searched the word "dangerous" in this output file and didn't find anything.

out.1482662 used the hand-written ccsd, and never restarted from a fort.55, so never a change in available RAM or "number of active orbitals", but also gave huge residuals.

I also repeated out.1482662 but with more RAM, but perhaps it's not so valuable to attach that output file because the results are the same, just without using the out-of-core algorithm to read from disk all the time.

I've now run a job with itol=18 and scftol=13 and and the SCF converged, but ovirt has still been on the second part of the alpha-alpha step for quite a while. If this doesn't work, perhaps I have to try with the CFOUR interface, but I have to wait until this job finishes since all the large RAM nodes are running other jobs at the moment.

With best wishes!!
Nike

#### File Attachment:

File Name: out.1462030.txt
File Size:15 KB

#### File Attachment:

File Name: out.1482423.txt
File Size:39 KB

#### File Attachment:

File Name: out.1482662.txt
File Size:23 KB