Problems converging TDDFT calculations

Find answers or ask questions regarding LSDalton calculations
Post Reply
AndreAnda
Posts: 6
Joined: 07 Aug 2015, 13:25
First name(s): André
Last name(s): Anda
Affiliation: University of Copenhagen
Country: Denmark

Problems converging TDDFT calculations

Post by AndreAnda » 07 Aug 2015, 13:49

I am calculating excitation energies using TDDFT but only half of my calculations are successful.

The input files differ only in small displacements of one dihedral angle and one H atom, but do not seem to have anything to do with the failed calculations.

The errors seem to be equally shared between (see attached .out files)

"SCF has failed to converge in 100 iterations" and

"Too many rejections; Number of rejections 7
This could be related to lack of integral accuracy or a complicated electronic
structure."

Increasing the integral accuracy did not help.

Any suggestions as to how to increase the success rate of the calculations will be greatly appreciated.

André Anda - University of Copenhagen
Attachments
2dTyrBchla_2dTyrBchla_0_0.6.out
(721.95 KiB) Downloaded 426 times
2dTyrBchla_2dTyrBchla_-4_0.3.out
(242.1 KiB) Downloaded 420 times

tkjaer
Posts: 300
Joined: 27 Aug 2013, 20:35
First name(s): Thomas
Last name(s): Kjaergaard

Re: Problems converging TDDFT calculations

Post by tkjaer » 07 Aug 2015, 14:27

It looks to be a complicated electronic structure

Concerning

2dTyrBchla_2dTyrBchla_0_0.6.out

1. there is a warning - you seem to only use 1 OpenMP thread but 4 MPI nodes - this is not optimal
2. it looks to converge smmothly, but slow. I suggest that you increase the maximum number of SCF iterations with

.
*DENSOPT
.MAXIT
300

Concerning

2dTyrBchla_2dTyrBchla_-4_0.3.out

I would try to use

.ARH FULL

instead of

.ARH

TK

AndreAnda
Posts: 6
Joined: 07 Aug 2015, 13:25
First name(s): André
Last name(s): Anda
Affiliation: University of Copenhagen
Country: Denmark

Re: Problems converging TDDFT calculations

Post by AndreAnda » 17 Aug 2015, 07:31

Both of them now worked. Thanks a lot!

André Anda

tkjaer
Posts: 300
Joined: 27 Aug 2013, 20:35
First name(s): Thomas
Last name(s): Kjaergaard

Re: Problems converging TDDFT calculations

Post by tkjaer » 17 Aug 2015, 08:16

Did the rate of success also increase?

Naturally you can do both for this.

*DENSOPT
.ARH FULL
.MAXIT
300

Let us know if we can be of further assistance

AndreAnda
Posts: 6
Joined: 07 Aug 2015, 13:25
First name(s): André
Last name(s): Anda
Affiliation: University of Copenhagen
Country: Denmark

Re: Problems converging TDDFT calculations

Post by AndreAnda » 03 Sep 2015, 12:49

Only 2 out of 99 calculations failed (due to "too many rejections..."). However, for a particular hydrogen displacement (0.6 Å, and not for 0.5 Å or 0.7 Å) it doesn't find the excited state corresponding to the optical transition we're looking for (the oscillator strengths are too weak). The ground state looks good. We use

.NEXCIT
5

but will increase this to see if that helps.

Thanks!
Attachments
2dTyrBchla_2dTyrBchla_6_0.6.out
(584.4 KiB) Downloaded 365 times

tkjaer
Posts: 300
Joined: 27 Aug 2013, 20:35
First name(s): Thomas
Last name(s): Kjaergaard

Re: Problems converging TDDFT calculations

Post by tkjaer » 04 Sep 2015, 10:05

I would be interested in figuring out why those 2 do not converge, so could you send the input?

Concerning you problems with excited state I agree that you should try to increase the number of excitation energies.

Remember that you can "RESTART" you calculation from the converged density of you previous calculation - so that you do not need to redo the SCF calculation.

*DENSOPT
.RESTART

You can also restart the calculation of excitation energies using

.RESTEXC
5

Note that (it does not look like it) if you system have a high degree of point group symmetry you may run into the problem that the particular point group symmetry was not included in the original set of trial vectors used for the excitation energies, which means it will not find the proper set of excitation energies.

TK

AndreAnda
Posts: 6
Joined: 07 Aug 2015, 13:25
First name(s): André
Last name(s): Anda
Affiliation: University of Copenhagen
Country: Denmark

Re: Problems converging TDDFT calculations

Post by AndreAnda » 04 Sep 2015, 13:03

Here are the output and input of the two failed calculations
Attachments
2dTyrBchla.dal
(144 Bytes) Downloaded 380 times
2dTyrBchla_2dTyrBchla_4_0.6.out
(531.63 KiB) Downloaded 400 times
2dTyrBchla_2dTyrBchla_-8_0.6.out
(397.91 KiB) Downloaded 388 times

tkjaer
Posts: 300
Joined: 27 Aug 2013, 20:35
First name(s): Thomas
Last name(s): Kjaergaard

Re: Problems converging TDDFT calculations

Post by tkjaer » 04 Sep 2015, 14:14

Thank you very much.

It looks like the calculations are very close to convergence (gradient = 2.67E-05 which means the energy is converged to around 1E-10 ) but fail to converge.

My thoughts would be that this is a DFT grid precision issue or an issue that needs a more careful investigation.

if you want to push these calculations through I suggest to use a "FINE" DFT grid.

*DFT INPUT
.FINE

TK

AndreAnda
Posts: 6
Joined: 07 Aug 2015, 13:25
First name(s): André
Last name(s): Anda
Affiliation: University of Copenhagen
Country: Denmark

Re: Problems converging TDDFT calculations

Post by AndreAnda » 22 Sep 2015, 08:12

The fine grid did squeeze them through. But the problem with excitation energies that do not fit in with the rest of the configuration space still remains.

Any thoughts on what's causing this?

Thanks again.

André Anda

tkjaer
Posts: 300
Joined: 27 Aug 2013, 20:35
First name(s): Thomas
Last name(s): Kjaergaard

Re: Problems converging TDDFT calculations

Post by tkjaer » 22 Sep 2015, 08:27

Concerning the excitation energies I can only suggest that you

1. Converge the SCF harder. However,

.CONVDYN
TIGHT

should be tight enough

2. request more excitation energies

3. If the SCF converge to a saddle point the excitation energies would not make sense - but this should not happen with ARH FULL - in principle you can check that you have converged to a minimum, but that is very expensive - as expensive as the full SCF.

can you send the output - maybe that will tell me more

AndreAnda
Posts: 6
Joined: 07 Aug 2015, 13:25
First name(s): André
Last name(s): Anda
Affiliation: University of Copenhagen
Country: Denmark

Re: Problems converging TDDFT calculations

Post by AndreAnda » 28 Sep 2015, 13:53

Here is an example of the output.

Can you learn anything from it?

André
Attachments
2dTyrBchla_2dTyrBchla_-8_0.6_fine.out
(444.08 KiB) Downloaded 364 times

luca
Posts: 7
Joined: 11 Mar 2014, 13:41
First name(s): Luca
Last name(s): De Vico
Affiliation: Copenhagen University
Country: Denmark

Re: Problems converging TDDFT calculations

Post by luca » 19 Aug 2016, 10:41

Dear LSDalton forum,
we are still having some difficulties with our TD-DFT calculations. While doing a geometry scan, that is moving one or very few atoms very very little at a time, we often encounters geometries where nothing works out as in those surrounding, creating strange and unexpected artifacts.
One of the most disturbing aspects is that computed excitation energies suddenly come out mixed up. Let me explain with an example. We are interested in the first excitation energy, and we compute 5 for good measure.

Here is the output of a "good" calculation:

Excitation Transition Dipole Moments Oscillator
Energies x y z Strengths
===============================================================================
0.05813173 3.21873928 -0.15735079 -0.39032106 0.40837123
0.07556057 0.58977443 1.32138545 0.04746374 0.10559066
0.08414462 -0.08317878 0.50510467 0.09980511 0.01525881
0.09505294 -0.00398517 -0.13386777 0.03063855 0.00119609
0.10261680 -0.32171268 -0.36315346 0.01186048 0.01611222

Here is the output from a "bad" one:

Excitation Transition Dipole Moments Oscillator
Energies x y z Strengths
===============================================================================
0.00426285 0.24270230 -0.39201328 0.79239283 0.00238852
0.04182679 -0.01733164 0.02886855 -0.06171046 0.00013780
0.05460040 0.75428362 -1.67398345 0.66631174 0.13887196
0.05769026 0.99870531 -2.18183275 0.18755275 0.22279912
0.06781730 -0.00222396 0.03686129 0.04607496 0.00015763
0.07504558 -0.02057032 -0.33897552 1.62795848 0.13836285
0.07601010 0.04845519 -0.06814299 -0.09954983 0.00085646
0.08729695 -0.04833878 0.09489194 0.01443687 0.00067216

As you can see, we increased the number of computed excitations to 8, to try and get better numbers, but to no avail. Even 12 did not help.
The third and forth reported excitations seem to be the one we want, but for some reason split in two, as also evinced by the oscillator strength. We tried to use the .DTHR keyword for *SOLVER, with a value of even 0.03 (I know, it should be an unnecessary exaggeration, but we tried everything) but it did not have any effect.
I attached the output files. I hope you can tell me what are we doing wrong, or what could we use to get those excitation energies to behave.
Thanks, regards
Luca De Vico
Attachments
tyr4_tyr_0_0.6.out
LSDalton output
(218.66 KiB) Downloaded 172 times
tyr_0_0.6.4040651.out
Standard output
(32.5 KiB) Downloaded 156 times

taylor
Posts: 519
Joined: 15 Oct 2013, 05:37
First name(s): Peter
Middle name(s): Robert
Last name(s): Taylor
Affiliation: Aarhus University
Country: Denmark

Re: Problems converging TDDFT calculations

Post by taylor » 19 Aug 2016, 13:06

I am not an LSDalton developer, nor even an expert user, although I have used the code quite a bit. But an excitation energy of 0.004 Eh is obviously ridiculous. One possibility is that this is a problem in Nature (that this particular geometry has some sort of resonance or near-resonant behaviour) that causes some sort of instability in the calculation. Another possibility is that the perceived numerical instability (which is clearly the issue) is intrinsic to some of the methodology as programmed (the algorithms or their implementation). A possible path forward in identifying problems might be to use a different method (this is to test the behaviour of the code, not as an alternative that would be better scientifically!). For example, if you take this problem geometry and run with a different functional, or even with straight SCF?

Best regards
Pete
P.S. Needless to say, I would defer to any of the LSDalton developers/experts who might post subsequently: this is not false modesty, I really am not an LSDalton expert!

luca
Posts: 7
Joined: 11 Mar 2014, 13:41
First name(s): Luca
Last name(s): De Vico
Affiliation: Copenhagen University
Country: Denmark

Re: Problems converging TDDFT calculations

Post by luca » 19 Aug 2016, 14:35

Hi Pete,
Thanks for your reply. We are producing a huge number of structures, and only a limited set of them shows this strange behavior. It would seem very un-physical if there was a "Nature" cause behind. One never knows, obviously, but ... in this case I'm pretty sure there is something else. We have so many single points that we are not collecting data by hand, we use scripts which check for a reasonable enough oscillator strength before collecting the corresponding energy. That's why it took us some time before we realized that what we thought were unfinished calculations were, actually, finished but with weird results. I will give it a try to another functional, but I don't want (at this point) redo the entire set of calculations with another method, it would require too much time. What we need is a way of "patching up" the holes in the surfaces we are,otherwise, producing.
Cheers!
/Luca

taylor
Posts: 519
Joined: 15 Oct 2013, 05:37
First name(s): Peter
Middle name(s): Robert
Last name(s): Taylor
Affiliation: Aarhus University
Country: Denmark

Re: Problems converging TDDFT calculations

Post by taylor » 19 Aug 2016, 15:25

No, I wasn't suggesting "redoing" anything, and certainly not "everything"! I was saying more that the more information that can be provided to the LSDalton mob, the more there's a chance they might be able to figure out that there's a problem. If you were to redo just a few calculations with some geometries that worked and one that went crazy, with (e.g.) a different functional,, or SCF, it might provide some information to someone seeking to find a bug and kill it. Conversely, it is essential to ensure that this isn't some sort of "freak of Nature" like an unexpected resonance (which might be an artefact of a particular functional, or basis set, for instance): it makes no sense for someone to chase a bug unless it's clear there is a bug, as opposed to "unsatisfactory results"...

Best regards
Pete

tkjaer
Posts: 300
Joined: 27 Aug 2013, 20:35
First name(s): Thomas
Last name(s): Kjaergaard

Re: Problems converging TDDFT calculations

Post by tkjaer » 19 Aug 2016, 19:20

As far as I remember the excitation energy of 0.004 Eh could indicate an issue with the SCF convergence - that the density is not fully converged

I suggest to use

.CONVDYN
TIGHT

when doing response in general

It does look like you have problems with the SCF convergence (converging very slowly - 76 iterations)

I suggest to use

**INTEGRALS
.THRESH
1.0D-15
**WAVE FUNCTIONS
.DFT
PBE0
*DFT INPUT
.FINE
*DENSOPT
.ARH FULL
.CONVDYN
TIGHT
.START
TRILEVEL
.NVEC
100
.MICROVECS
40
.MAXIT
300
**RESPONS
.NEXCIT
1
*SOLVER
.NSTART
10
.CONVDYN
TIGHT
*END OF INPUT

this will only converge the first excited state (save time) but use 10 different startvectors (or some other value depending on molecule symmetry) that hopefully contain the proper symmetry to ensure the lowest excited state is the right one.

luca
Posts: 7
Joined: 11 Mar 2014, 13:41
First name(s): Luca
Last name(s): De Vico
Affiliation: Copenhagen University
Country: Denmark

Re: Problems converging TDDFT calculations

Post by luca » 21 Aug 2016, 19:20

The SCF convergence was another issue, that we tried to solve previously. The choice of

.ARH DAVID
.CONVDYN
STANDARD

is the one that ensures we always get an SCF to converge, even if slowly. We found it after many try and errors. However, it may very well be that it is insufficient. I will try one thing at the time. First the suggested 1 excitation with 10 starting vectors. If that were not enough I will try also to change the SCF convergence, but it might create other problems.

Cheers!
/Luca

Post Reply

Who is online

Users browsing this forum: No registered users and 1 guest