dft convergence problem

 Posts: 22
 Joined: 15 Dec 2013, 22:25
 First name(s): Angelo
 Middle name(s): Domenico
 Last name(s): Quartarolo
 Affiliation: High School
 Country: Italy
dft convergence problem
Dear Dalton users,
after 60 DIIS and 60 QCDFT iterations, my wavefunction didn't converge.It is a palladiumphthalocyanine complex. I rerun the calculation switching off DIIS and the energy is slowly converging. My question is about which parameters (apart increasing the number of iterations) I have to use for DIIS wavefunction convergence and how to exactly set them. I know something about damping factor or level shift but I never applied them in Dalton.
I attached the input and output files
Thanks in advance
Quartarolo Domenico
after 60 DIIS and 60 QCDFT iterations, my wavefunction didn't converge.It is a palladiumphthalocyanine complex. I rerun the calculation switching off DIIS and the energy is slowly converging. My question is about which parameters (apart increasing the number of iterations) I have to use for DIIS wavefunction convergence and how to exactly set them. I know something about damping factor or level shift but I never applied them in Dalton.
I attached the input and output files
Thanks in advance
Quartarolo Domenico
 Attachments

 PdPhthalocyanine.mol
 (6.43 KiB) Downloaded 447 times

 PdPhthalocyanine.dal
 (229 Bytes) Downloaded 454 times

 PdPhthalocyanine.out
 (267.06 KiB) Downloaded 395 times

 Posts: 379
 Joined: 27 Jun 2013, 18:44
 First name(s): Hans Jørgen
 Middle name(s): Aagaard
 Last name(s): Jensen
 Affiliation: Universith of Southern Denmark
 Country: Denmark
Re: dft convergence problem
Could you upload the output when running without DIIS ?
(Look at the Mulliken charges on Pd, it is very clear that your calculation starts oscillating terribly. This should not happen with second order convergence.)
(Look at the Mulliken charges on Pd, it is very clear that your calculation starts oscillating terribly. This should not happen with second order convergence.)

 Posts: 379
 Joined: 27 Jun 2013, 18:44
 First name(s): Hans Jørgen
 Middle name(s): Aagaard
 Last name(s): Jensen
 Affiliation: Universith of Southern Denmark
 Country: Denmark
Re: dft convergence problem
PS. and I suggest you switch to Dalton2015

 Posts: 22
 Joined: 15 Dec 2013, 22:25
 First name(s): Angelo
 Middle name(s): Domenico
 Last name(s): Quartarolo
 Affiliation: High School
 Country: Italy
Re: dft convergence problem
I attached the temporary output file (NODIIS).Now I'm trying Dalton2015.
thanks
thanks
 Attachments

 DALTON.OUT
 (81.43 KiB) Downloaded 441 times

 Posts: 379
 Joined: 27 Jun 2013, 18:44
 First name(s): Hans Jørgen
 Middle name(s): Aagaard
 Last name(s): Jensen
 Affiliation: Universith of Southern Denmark
 Country: Denmark
Re: dft convergence problem
The NODIIS output looks very good  it is converging smoothly, and to sensible atomic charges. Did you run out of time or ?

 Posts: 595
 Joined: 15 Oct 2013, 05:37
 First name(s): Peter
 Middle name(s): Robert
 Last name(s): Taylor
 Affiliation: Tianjin University
 Country: China
Re: dft convergence problem
In addition to Hans Joergen's cogent suggestions, can I add something from a different perspective? You say your compound is a Pdphthalocyanine: obviously from looking at the input it is some sort of substituted form because it has a bunch of oxygens and sulfurs and so forth. How did you generate the input (the .mol file, I mean)? I ask because if you look at the principal moments of inertia, they seem to be extraordinarily close to those of a planar symmetric top. That is, in terms of principal moments I_{A}, I_{B}, and I_{C}, you are very close to having I_{A} = I_{B} (i.e., symmetric top  molecule with a higher than twofold principal axis of rotation) and I_{A} + I_{B} = I_{C}, which would imply planarity perpendicular to the principal axis.
Do you expect the system you are studying to have these symmetry properties, at least ideally? Perhaps (given at least the symmetry of the bare phthalocyanine ligand) D_{4h}? If your structure is from e.g. Xray diffraction it may be that the coordinates are "noisy" enough that they do not display the exact symmetry, but this might be a consequence of the experimental structure refinement. If your structure is from another program, such as a builder or crudelevel (like molecular mechanics) optimization, perhaps this has not produced the desired or expected symmetry?
I ask this because while one might hope (and I have posted this elsewhere before and am not prepared to post it again) that turning off symmetry will obviate the need to worry about orbital occupations across symmetries and how to estimate them and which states to optimize and so forth, in a system with symmetry this is a tactical error. For example, for a symmetric top there will be degenerate orbitals, and if the program occupies degenerate orbitals incompletely there will be all sorts of symmetrybreaking issues in a wave function calculation that can trip things up. (I do not suggest this is the cause of the behaviour in the DIIS calculation, which is just insane, with iterations producing positive total energies!) But I do suggest that if, naively, your molecule could have some symmetry, and particularly higher symmetry than D_{2h}, you would do well to (a) symmetrize your input structure so that it actually has this symmetry, and then either leave symmetry to the program (take out Nosymmetry) or dictate the symmetry yourself. See my posting at the top of "running Dalton" for assistance with the latter.
Best regards
Pete
Do you expect the system you are studying to have these symmetry properties, at least ideally? Perhaps (given at least the symmetry of the bare phthalocyanine ligand) D_{4h}? If your structure is from e.g. Xray diffraction it may be that the coordinates are "noisy" enough that they do not display the exact symmetry, but this might be a consequence of the experimental structure refinement. If your structure is from another program, such as a builder or crudelevel (like molecular mechanics) optimization, perhaps this has not produced the desired or expected symmetry?
I ask this because while one might hope (and I have posted this elsewhere before and am not prepared to post it again) that turning off symmetry will obviate the need to worry about orbital occupations across symmetries and how to estimate them and which states to optimize and so forth, in a system with symmetry this is a tactical error. For example, for a symmetric top there will be degenerate orbitals, and if the program occupies degenerate orbitals incompletely there will be all sorts of symmetrybreaking issues in a wave function calculation that can trip things up. (I do not suggest this is the cause of the behaviour in the DIIS calculation, which is just insane, with iterations producing positive total energies!) But I do suggest that if, naively, your molecule could have some symmetry, and particularly higher symmetry than D_{2h}, you would do well to (a) symmetrize your input structure so that it actually has this symmetry, and then either leave symmetry to the program (take out Nosymmetry) or dictate the symmetry yourself. See my posting at the top of "running Dalton" for assistance with the latter.
Best regards
Pete

 Posts: 22
 Joined: 15 Dec 2013, 22:25
 First name(s): Angelo
 Middle name(s): Domenico
 Last name(s): Quartarolo
 Affiliation: High School
 Country: Italy
Re: dft convergence problem
The NODIIS calculation,which is converging, is still running.hjaaj wrote:The NODIIS output looks very good  it is converging smoothly, and to sensible atomic charges. Did you run out of time or ?
The calculation running with Dalton2015 and DIIS algorithm is not converging.(attached file)
About the remarks made by prof. Taylor:
a) the molecule is peripherically substitued by eight methyl sulphonate (CH3SO2) groups,the system was optimized without symmetry
at M06/631G(d) level of theory and Stuttgart pseudopotential for Pd atom (another software). The molecular point group is D4 . I'm not sure but can Dalton handle these point groups? (it seems no from pag. 173175 of the manual).
b) Single point calculation with DIIS at the same level and another software converged.This is also why asked about dynamical damping factor or others to be set in the input file, other tricks like choosing different initial guess density (H1DIAG) or fine grid don't work.
Thanks
domenico
 Attachments

 DALTON2015.out
 (82.39 KiB) Downloaded 381 times

 Posts: 595
 Joined: 15 Oct 2013, 05:37
 First name(s): Peter
 Middle name(s): Robert
 Last name(s): Taylor
 Affiliation: Tianjin University
 Country: China
Re: dft convergence problem
Dalton cannot handle D_{4} in its full form, but it can handle the subgroup D_{2}. Although your molecular geometry is close to D_{4}, it is not exactly D_{4}. This is something to be aware of because in my experience these "close but no cigar" situations can give misleading results. Not long ago I was looking at a geometry generated by someone who had run without symmetry and which looked close (in this case) to C_{3v} symmetry, but wasn't quite, being recognized only as C_{1}. I symmetrized the geometry to have exact C_{3v} symmetry, reoptimized within C_{3v}, and found a minimum (established as such by calculating the frequencies) at a slightly lower energy than that obtained in C_{1}. So it pays to check these things.
Best regards
Pete
Best regards
Pete

 Posts: 22
 Joined: 15 Dec 2013, 22:25
 First name(s): Angelo
 Middle name(s): Domenico
 Last name(s): Quartarolo
 Affiliation: High School
 Country: Italy
Re: dft convergence problem
I checked D4 against C1 geometry, the C1 energy is more stable by only 0.1 kcal/mol.
After trying different ways for DIIS convergence, I found a solution from level shift setting (.SHIFT 0.1
),it saved about two days of calculation in comparison to the QC algorithm (same initial conditions).
I just used level shift also for the corresponding platinum complex and it is also well going to convergence.
best regards
domenico
After trying different ways for DIIS convergence, I found a solution from level shift setting (.SHIFT 0.1
),it saved about two days of calculation in comparison to the QC algorithm (same initial conditions).
I just used level shift also for the corresponding platinum complex and it is also well going to convergence.
best regards
domenico
 Attachments

 DALTONshift.out
 (209.37 KiB) Downloaded 399 times

 Posts: 379
 Joined: 27 Jun 2013, 18:44
 First name(s): Hans Jørgen
 Middle name(s): Aagaard
 Last name(s): Jensen
 Affiliation: Universith of Southern Denmark
 Country: Denmark
Re: dft convergence problem
Would you also upload the final NODIIS output, please? I want to analyze it, why it took so long time.
 Hans Jørgen
 Hans Jørgen

 Posts: 22
 Joined: 15 Dec 2013, 22:25
 First name(s): Angelo
 Middle name(s): Domenico
 Last name(s): Quartarolo
 Affiliation: High School
 Country: Italy
Re: dft convergence problem
Dear Prof. Jørgen
here it is the output you asked.
best regards
domenico
here it is the output you asked.
best regards
domenico
 Attachments

 DALTON_NODIIS.out
 (146.75 KiB) Downloaded 393 times

 Posts: 14
 Joined: 12 Dec 2013, 12:43
 First name(s): Erik
 Middle name(s): Donovan
 Last name(s): Hedegård
 Affiliation: University of Southern Denmark
 Country: Denmark
 Contact:
Re: dft convergence problem
Hi
Be aware that the NODIIS option as far as i know is not implemented for DFT in DALTON, and this is probably why the calculation does not converge.
I would suggest to converge with NODIIS for hartree fock (.HF instead of .DFT) and restart the calculation from the converged HF solution. The restart can be done by inserting
*ORBITAL INPUT
.MOSTART
NEWORB
Erik
Be aware that the NODIIS option as far as i know is not implemented for DFT in DALTON, and this is probably why the calculation does not converge.
I would suggest to converge with NODIIS for hartree fock (.HF instead of .DFT) and restart the calculation from the converged HF solution. The restart can be done by inserting
*ORBITAL INPUT
.MOSTART
NEWORB
Erik

 Posts: 22
 Joined: 15 Dec 2013, 22:25
 First name(s): Angelo
 Middle name(s): Domenico
 Last name(s): Quartarolo
 Affiliation: High School
 Country: Italy
Re: dft convergence problem
Hi,
the time comparison was really not correct, infact it was made between dalton2013 for NODIIS and dalton2015 for DIIS + shift level. So I tried to calculate all with dalton2015, but the NODIIS case now exits with: QCSCF not implemented/tested for DFT.
Anyway the energy calculated with NODIIS (dalton2013) seems "correct", that is, it matches that calculated with another popular software.
the time comparison was really not correct, infact it was made between dalton2013 for NODIIS and dalton2015 for DIIS + shift level. So I tried to calculate all with dalton2015, but the NODIIS case now exits with: QCSCF not implemented/tested for DFT.
Anyway the energy calculated with NODIIS (dalton2013) seems "correct", that is, it matches that calculated with another popular software.

 Posts: 379
 Joined: 27 Jun 2013, 18:44
 First name(s): Hans Jørgen
 Middle name(s): Aagaard
 Last name(s): Jensen
 Affiliation: Universith of Southern Denmark
 Country: Denmark
Re: dft convergence problem
Well, Erik, this one definitely did not have problems. From the output uploaded by Domenico:
and as you can see from the GRDNRM (gradient norm) column, there is basically quadratic convergence, but it uses many more Fock matrix constructions (basically each micro iteration costs a Fock matrix constructions, so each macroiteration costs (1+ITMIC) Fock matrix constructions. This adds up to 72 Fock matrix constructions compared to the 19 needed for your DIIS with levelshift calculations, a factor of 3.8. The factor between total CPU time in SIRIUS for the two calculations is 3.6, so it fits.
The quadratically convergent secondorder optimization is very stable, but DIIS is definitely a lot faster (if it converges ).
Code: Select all
ITER ITMIC EMCSCF GRDNRM RATIO STPLNG

1 2 6453.925490820438 95.7650991595 0.000000 0.7354361532
2 3 6486.736500416813 12.5027795305 0.947071 0.7301374313
3 3 6493.035211768460 6.2110179258 1.006403 0.7450703036
4 12 6495.755920070926 1.8206305618 1.014840 0.4796406532
5 3 6496.107510508631 0.4792035008 1.061830 0.1470590304
6 12 6496.128679426918 0.0867574840 1.065402 0.0271471983
7 12 6496.129363474109 0.0105940444 1.024387 0.0027834898
8 12 6496.129372944150 0.0005076738 1.011071 0.0001561808
9 3 6496.129372972169 0.0000450173 1.000000 0.0000155571
10 0 6496.129372973119 0.0000071568 1.000000 0.0000000000
The quadratically convergent secondorder optimization is very stable, but DIIS is definitely a lot faster (if it converges ).

 Posts: 22
 Joined: 15 Dec 2013, 22:25
 First name(s): Angelo
 Middle name(s): Domenico
 Last name(s): Quartarolo
 Affiliation: High School
 Country: Italy
Re: dft convergence problem
Thanks for the clarification
Who is online
Users browsing this forum: Google [Bot] and 8 guests