Author Topic: Excited state optimization  (Read 12471 times)

Beverly

  • Newbie
  • *
  • Posts: 8
  • Karma: +0/-0
Excited state optimization
« on: May 17, 2010, 07:30:49 PM »
Hi everyone,
  i was wondering if someone could help me out. I'm attempting to optimise the fourth excited triplet of a small molecule (around 33 atoms) I've been using RPAT with conv=8 and when i call jobex I have called jobex -ri -ex -gcart 6 -energy 6.
  My ground state structure has been optimized with SV(p) and this is the basis set i have used for my excited state calculations. The problem I am having is my calculation simply won't optmize. |dE/dxyz| constantly oscillates at around 0.06-0.08
  I tried initially using default settings for conv and gcart but with the same result.
 
  If anyone has any ideas on how i could fix this it would be much appreciated,

Thanks
Beverly

jbaltrus

  • Full Member
  • ***
  • Posts: 71
  • Karma: +0/-0
Re: Excited state optimization
« Reply #1 on: May 19, 2010, 07:41:33 PM »
Beverly,

exactly the same problem here for most of the structures. I've been told to get rid of hessaprox, let me know if you came up with any solutions for this problem

christof.haettig

  • Global Moderator
  • Sr. Member
  • *****
  • Posts: 291
  • Karma: +0/-0
    • Hattig's Group at the RUB
Re: Excited state optimization
« Reply #2 on: July 20, 2010, 05:35:30 PM »
Geometry optimizations on higher excited states (i.e. not S1 or T1) can become tricky, since one often runs into regions of avoided crossings or conical intersections or where accidentally two states get very close in energy. In these regions it can easily happen that your calculation converges to a different state or jumps between two excited states.
Here, you really have to check the electronic structure of the excited state of interest and those close by  and make sure
that at the points in these regions the gradient in computed for the right state.

Christof

jbaltrus

  • Full Member
  • ***
  • Posts: 71
  • Karma: +0/-0
Re: Excited state optimization
« Reply #3 on: July 23, 2010, 03:08:21 AM »
I actually think that TDDFT or CIS are very very system dependent and very unfortunately it's almost a coincidence if a certain molecule will undergo geometry optimization with -ex in turbomole and most likely in other packages. I am pretty sure what happened to Beverly, e.g. in gradient part Davidson iterations don;t even think of converging and eventually one runs out of iterations. That probably is conical crossection or excited state comes close to gound state or something else that was mentioned in the email above.

This is very frustrating as I work with series of similar molecules, some of them I can optimize in excited state tddft or cis and some all of a sudden fail. So my project stale because of that and there is not clear path to go. I tried ricc but that's just unsustainable for large molecules.

I was wondering if any of turbomole developers could comment on a situation like that, e.g. there is a lot of knowledge why tddft or cis could fail, but there is not many suggestions what can be done or which route one should go to -ex optimize a ~60 atom molecule

Jonas

christof.haettig

  • Global Moderator
  • Sr. Member
  • *****
  • Posts: 291
  • Karma: +0/-0
    • Hattig's Group at the RUB
Re: Excited state optimization
« Reply #4 on: July 23, 2010, 02:42:31 PM »
There are lot of suggestions around what should be checked and what could be done. But what is appropriate depends very much on the situation, the method and the implementation...

  * if the system gets close to a conical intersection with the ground state can be detected by unusual small excitation energies and slow convergence for the ground state SCF, CC2 or DFT calculation. I would say, whenever the lowest excitation energy drops below 1 eV, one must doubt that single reference methods are appropriate and should check if doesn't face a serious multireference case for ground state....

  * slow convergence in the calculation for the excitation energies is usually due to either a very poor start  guess or near degeneracies between excited states of the same symmetry. In the later case one should try to include all close lying excited states (since there eigenvectors usually mix a lot). During geometry optimizations one can detect such situations usually either by small gaps in the excitation energies (f.x. in previous iterations) or strong changes in the most important contributions to the eigenvectors (printed in the output). Also root flipping due to jumps over avoided intersections can be detected this way. In that case it is often helpful to compute some intermediate points between the last structure before and the first structure after root flipping to check what is going and what is a good point to restart the geometry optimization. Sometimes it helps to restrict the max. steps size for the geometry optimization.

Clearly, there is not 'one recipe'. It's true research. Trial and error.

Christof

jbaltrus

  • Full Member
  • ***
  • Posts: 71
  • Karma: +0/-0
Re: Excited state optimization
« Reply #5 on: July 24, 2010, 06:29:00 AM »
Christof,

that is very useful. Maybe you can help me with a specific example?

at some point in excited state optimization I am getting this during egrad typical of my failing systems:

                      Block Davidson iteration


 total number of roots to be determined:   4


 maximum core memory set to   200 MB,
 corresponding to       22 vectors in CAO basis


 maximum number of simultaneously treated vectors (including degeneracy):   4


 Iteration IRREP Converged      Max. Euclidean
                 roots          residual norm

    1       a        0        3.337522529394235D+00
 
    2       a        0        2.854556752014250D+00
 
    3       a        0        2.817250468160716D+00
 
    4       a        0        2.711339667773352D+00
 
    5       a        0        2.770424474706550D+00
 
    6       a        0        2.765030833642312D+00
 
    7       a        0        2.758343394180049D+00
 
    8       a        0        2.754071221271158D+00
 
    9       a        0        2.763318999565214D+00
 
   10       a        0        2.764305534461768D+00
 
   11       a        0        2.759222813723960D+00
 
   12       a        0        2.759320774710940D+00
 
   13       a        0        2.762143311483079D+00
 
   14       a        0        2.763551531945221D+00
 
   15       a        0        2.764545275537617D+00
 
   16       a        0        2.763505517975601D+00
 
   17       a        0        2.763316324763632D+00
 
   18       a        0        2.764235300936229D+00
 
   19       a        0        2.763170339338126D+00
 
   20       a        0        2.762566707118540D+00
 
   21       a        0        2.762374488574519D+00
 
   22       a        0        2.762356627722698D+00
 
   23       a        0        2.762367111239716D+00
 
   24       a        0        2.762335675116838D+00
 
   25       a        0        2.762317572823165D+00
 


 Warning! No convergence within  25 iterations.
 Unless you have specified a very low $escfiterlimit,
 this is a reason to worry!

When I look at roots they don;t seem to be telling me anything:

 IRREP   Vector     Eigenvalue           Euclidean residual norm

  a         1   -6.410929294200665D-01    2.762317572823165D+00
            2   -4.907372544150190D-02    7.799946539795827D-01
            3   -3.088690544985784D-02    1.242132979626735D+00
            4   -3.179104291302082D-04    5.015595530551463D-02

 Which case is it? Which way should I go in my trial and error search?

I am running

$dft
   functional b-p
   gridsize   m3
$scfinstab rpas
$soes
a            4
$spectra nm
$cdspectra nm
$exopt 1
$scfconv   7
$lastdiag

Jonas