Convergence Issue

38 views
Skip to first unread message

Michael Chen

unread,
Aug 4, 2026, 12:15:48 PMAug 4
to dirac-users
Hello DIRAC experts,

I'm relatively new to DIRAC, but I have started encountering convergence issues while performing a potential energy curve scan. What's the best way to help with the convergence?

I've attached an example from one of the output files with this issue.

Best,
Michael
uo_ecp4_distance=1.65.out

peterso...@gmail.com

unread,
Aug 4, 2026, 12:28:30 PMAug 4
to dirac...@googlegroups.com
Hi Michael,

I’m interested in what others have to say, but I think the default convergence criterion is just a bit too tight - it’s converged to better than 1e-12 on the energy. 

regards,  -Kirk
--
You received this message because you are subscribed to the Google Groups "dirac-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to dirac-users...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/dirac-users/104fb17c-6a0a-49bf-a5ff-3912a00ed9f5n%40googlegroups.com.

Hans Jørgen Aagaard Jensen

unread,
Aug 4, 2026, 2:48:31 PMAug 4
to dirac...@googlegroups.com
Hi Michael and Kirk,

It is NOT the default convergence criterion; 1.0e-9 was specified in input. That is sharper than the 15 digit numerical accuracy of double precision floating point numbers of the computerscan handle; in fact you can see from the output that after SCF is converged below 1.e-7 there is no further progress in convergence because of numerical noise because of the “15 digits”.

Wrt to the energy, there are 3 digits in front of the period. thus with with 12 decimal places one has reached the 15 digit limit - this is the reason starts to oscillate when converged to 1.e-12.

As SCF energy is quadratic in convergence threshold, you have converged to 1.e-12 when the gradient has converged to 1.e-6, as you can see in output. 

If you want microhartree accuracy, then a convergence threshold of 5.e-4 is enough.

It.  166    -548.4758144589      4.52D-09 -2.74D-05  4.16D-06   DIIS   9   12.17390000s   LL             Tue Aug  4
It.  167    -548.4758144589      8.83D-11 -2.92D-06  2.97D-06   DIIS   9   12.08310000s   LL             Tue Aug  4
It.  168    -548.4758144589      1.14D-13  4.41D-07  2.59D-06   DIIS   9   11.88200000s   LL             Tue Aug  4
It.  169    -548.4758144589     -2.41D-11  1.02D-06  2.15D-06   DIIS   9   11.95720000s   LL             Tue Aug  4
It.  170    -548.4758144589     -1.67D-11  1.01D-06  1.73D-06   DIIS   9   11.92760000s   LL             Tue Aug  4
It.  171    -548.4758144589     -1.19D-11  1.09D-06  1.41D-06   DIIS   9   11.94750000s   LL             Tue Aug  4
It.  172    -548.4758144589     -4.77D-12  5.01D-07  1.33D-06   DIIS   9   11.88190000s   LL             Tue Aug  4

Any convergence below 1.e-6 is irrelevant for the energy and simply waste of time.

I hope this helps.

Another thing is that the initial guess for the occupied orbitals is clearly terrible, and this leads to the many iterations.
I speculate if the poor initial guess i connected to the use of RECPs, I think the starting initial density guess is for all electron calculations and that is the reason it is so poor. The Dirac authors have themselves done very few RECP calculations, so we have prioritized improving other things in Dirac. Dirac is open-source, so anyone interested in  creating a better initial guess for RECP calculations are very welcome to contribute!

Best regards, Hans Jørgen.

peterso...@gmail.com

unread,
Aug 4, 2026, 3:00:38 PMAug 4
to dirac...@googlegroups.com
Hi Han Jørgen,

aha, you’re exactly right!   I didn’t look the input in this regard.  Thanks for clarifying!

best regards,

-Kirk

Chen, Michael

unread,
Aug 4, 2026, 3:35:01 PMAug 4
to dirac...@googlegroups.com
Hi Hans Jørgen and Kirk,

Two points:
  1. I specified 1.0D-9 for .EVCCNV because the manual specified for correlated methods that is what it should be set to. Are you suggesting for this to be decreased even for correlated methods like CCSD(T)?
  2. I realized after looking through the output file that the MO coefficients from the previous checkpoint was not used even though the checkpoint file appears to be present. I thought it will read from the checkpoint file if it is available? Is there something wrong with the checkpoint file?

    *** INFO *** No trial vectors found. Using sum of fitted atomic potentials as start potential. 
Best,
Michael

From: dirac...@googlegroups.com <dirac...@googlegroups.com> on behalf of peterso...@gmail.com <peterso...@gmail.com>
Sent: Tuesday, August 4, 2026 2:00 PM
To: dirac...@googlegroups.com <dirac...@googlegroups.com>
Subject: Re: [dirac-users] Convergence Issue
 
You received this message because you are subscribed to a topic in the Google Groups "dirac-users" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/dirac-users/WhDKYfqUB04/unsubscribe.
To unsubscribe from this group and all its topics, send an email to dirac-users...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/dirac-users/62E79527-70CD-C14E-BB02-088CA6DE82C1%40hxcore.ol.

Hans Jørgen Aagaard Jensen

unread,
Aug 4, 2026, 4:06:10 PMAug 4
to dirac...@googlegroups.com, Luuk Visscher
Hi again,

  1. I do not know who put that in the manual. For CCSD(T) the energy accuracy should be linear in the SCF convergence, thus 1.e-6 is enough for microhartree in CCSD(T). Maybe RECP has more numerical noise than 4c or x2c ...
  2. I am not sure the checkpoint file works for RECP ... Lucas ??

Best, Hans Jørgen.

Fra: dirac...@googlegroups.com <dirac...@googlegroups.com> på vegne af Chen, Michael <mc...@illinois.edu>
Dato: tirsdag, 4. august 2026 kl. 21.35
Til: dirac...@googlegroups.com <dirac...@googlegroups.com>
Emne: Re: [dirac-users] Convergence Issue

Michael Chen

unread,
Aug 4, 2026, 4:13:25 PMAug 4
to dirac-users
I realized there was a typo that caused an error when reading the checkpoint file. 

I can change .EVCCNV back to 1.0D-6 if that is sufficient enough for CCSD(T).

Best,
Michael 

Miroslav Iliaš

unread,
Aug 4, 2026, 5:13:49 PMAug 4
to dirac...@googlegroups.com

Michael Chen

unread,
Aug 5, 2026, 4:27:20 PMAug 5
to dirac-users
Hello DIRAC experts,

I now have issues with getting the CCSD calculation to converge. It seems to oscillate quite a bit. I've tried changing different parameters from the documentation, but nothing has worked so far.

Is there something I should modify to help it converge? Here is the output file.

Best,
Michael

uo_ecp5_distance=2.70.out

peterso...@gmail.com

unread,
Aug 5, 2026, 5:03:55 PMAug 5
to dirac...@googlegroups.com
Hi Michael,

if you look at the output of your .RESOLVE calculation, you’ll see that the ground state is starting to get pretty multireference at this rather long bond length. I’m guessing that has something to do with it.

best,  -Kirk

   -548.41746800

                        2  10010110                            0.6100    0.7810    0.0000

                        3  10100101                            0.3899   -0.6244    0.0000


From: dirac...@googlegroups.com <dirac...@googlegroups.com> on behalf of Michael Chen <mc...@illinois.edu>
Date: Wednesday, August 5, 2026 at 1:27 PM
To: dirac-users <dirac...@googlegroups.com>
Subject: Re: [dirac-users] Convergence Issue

Reply all
Reply to author
Forward
0 new messages