Summary of convergence problem with TDDFT in G98

From: sergiusz kwasniewski <sergiusz.kwasniewski \\at// luc.ac.be>
Subject: Summary of convergence problem with TDDFT in G98
Date: Wed, 5 Jan 2000 14:57:15 +0100 (MET)
Hi all,
 Thanks for the many responses I got. I did not try any of the propositions
 yet, because of a temporary lack of cpu time. I'll keep you informed if
 I still can't manage to run the calculations...
 I've read about level shift in some CASPT2 calculations ... But if I'm not
 mistaken, this should be handled with care, and you would need some
 experience in it in order to be able to apply it in a correct way.
 What is it about actually ? Are there some nice references that cover
 this shift ?
 As mentioned before I would summarize all the answers that I would get
 to the following problem:
 (BTW: TD-DFT stands for time-dependent density functional theory)
 <
 I've been trying a TD-DFT calculation (SP) with Gaussian98 and I've got
 the following message:
 <<
  Requested convergence on RMS density matrix=1.00D-08 within  64 cycles.
  Requested convergence on MAX density matrix=1.00D-06.
  Restarting incremental Fock formation.
  Restarting incremental Fock formation.
  Restarting incremental Fock formation.
  >>>>>>>>>> Convergence criterion not met.
  SCF Done:  E(RB+HF-LYP) =  -543.828916188     A.U. after   65 cycles
              Convg  =    0.3546D+01             -V/T =  2.0021
              S**2   =   0.0000
  Convergence failure -- run terminated.
  Error termination via Lnk1e in /usr/local/bin/g98.a7/l502.exe
 >>
 What keyword is best used to be able to solve this problem ?
 >Maxcycle ? Perhaps it's a problem with the grid ? All suggestions are
 >welcome. I'll summarize.
 >
 **********
 Hello serge,
 Obviously your criteria of convergence (convg) is not good at all, even
 after 65 cycles. Probably that the SCF procedure is oscillating. If it is
 so, it is useless to increase the number of cycles with the keyword
 "maxcyc". You could check it by introducing the keyword "p"
 in the command
 line and see how the SCF procedure behaves at each cycle. The complete
 command is: #p RHF/basis B3LYP ...
 Then, if the SCF procedure oscillates, you could try to apply a level shift.
 This is often usefull for large systems.
 The syntax is explained in the keyword "SCF".
 Hope this help,
 best wishes
 Pascal Boulet <Pascal.Boulet \\at// chiphy.unige.ch>
 **********
 Hi,
 to help you i need your input-line. These are just guesses:
 MAXCYCLE may be good, but sometimes even this wont work, then there could
 be something wrong with your starting geometry.
 by the way: what is a TD-DFT-Calculation??
 hope you get along
 tim Jödicke <jodickt \\at// uni-muenster.de>
 **********
 Dear Serge:
   I will use scf=qc (sometimes include maxcycle=XXXX XXXX: a larger
 number), although it takes me long time, it works. The usual thought is
 using scf=v-shift, but  for non-degeneracy (non-near-degeneracy) and with
 large LUMO-HOMO gap, it's better to use qc.
   This problem occurs often for pure DFT( I mean no exact exchange mixed).
  Hope it helpful.
  Wei Quan Tian <wtian \\at// uoguelph.ca>
 **********
 Dear Sergiusz,
    I don't think that keywords like qc or am will help you. These errors
 occur mostly because of the stupid Hueckel guess. I would try to start
 >from a smaller unpolarized BS without any diffuse functions (and
 scf=(maxcycle=900) ). Afterwards you can read the guess from that
 calculation.
    If you are dealing with transition metals or huge pi-systems and a
 very small HOMO-LUMO gap, the vshift command might help.
 I hope that will help!
 Ciao,
            Alexander Wittkopp  <awittko \\at// gwdg.de>
 **********
 Dear;
 Please use the iop option of iop(5/13=1) Although you fail the convergence
 of SCF uou can keep your calculation for TD-DFT and finally get some
 stuffs for excitations. Otherwise you can increase SCF cycle!
 Good luck!
 Ohyun Kwon <kwonohy \\at// auburn.edu>
 **********
   Dr. Kwasniewski,
    The information here is minimal so the analysis must of course be
 limited but given the fact that the convergence criteria is down only
 to 3.5 I highly doubt that increasing cycles will have any positive
 effect.
    You don't state which basis set you are trying or any details on
 the chemistry but I suspect you have a metal or some other system with
 low lying virtual and it may well be oscillating.  Further you likely
 have an extended basis set which gives the SCF more degrees of freedom
 to flounder in.
    My best suggestion would be a series of calculations naming the checkpoint
 file and the recovering SCF information at each step.
    1) #P B3LYP/STO-3g single point and look at the eigenvalues and make
       sure you have a HOMO/LUMO gap, best if eigenvalues change sign and
       the gap is 0.2 or greater.
    2) If small or inverted HOMO/LUMO gap try # B3LYP/STO-3G Guess=Read
       STABLE=OPT to confirm you got the true ground state before
       continuing.
    3) Once you are convinced you have the ground state use
       # B3LYP/6-31G(d) Guess=Read SCF=Tight or some other intermediate
       basis set with tighter SCF convergence.  If the HOMO/LUMO gap is
       still small consider SCF=(Tight,VShift=300) to help control mixing.
    4) Finally shift to your extended basis set.  If you have a lot of
       diffuse functions consider adding SCF=(NoVarAcc,Tight) along with
       GUESS=Read.  You can add TD at this point as well.
   If this is still not successful please let me know and include more
 details of the system.
 (Gaussian help <gaussian.com!help%gaussian.com \\at// gaussian.com>)
 **********
 The optimal strategy depends on what actually happens.
 (1) The energy is almost constant (changing by 1E-6 a.u.). This might
     indicate you need a better integration grid
 (2) The energy osscillates. --> Try larger vshift
 (3) The energy is monotonic, but converges slowly --> Increase MaxCycle
 (4) Nothing helps. Run a HF calculation, and use the HF orbitals from
     the checkpoint file as starting orbitals in DFT
 <Christoph.van.Wuellen \\at// ruhr-uni-bochum.de>
 **********
 Hi Serge,
 probably you have a bad starting geometry. So you have two choices:
 first: Maybe the convergence will be reached within more than 64 cycles, so try
 the keyword "SCF(MaxCycle=200)"
 second: you can try to improve the geometry while reducing the cutoff values for
 the SCF procedure. So I would try the combination of keywords
 "SCF=Sleazy" and
 "Opt(MaxCycle=3)". Then you perform a geometry optimization with
 reduced cutoff
 values which takes 3 cycles and will provide you an improved geometry for
 further
 calculations.
 Bye
 Thomas Mehnert <tmehnert \\at// physik.tu-muenchen.de>
 		     _________________________________________________
 			Sergiusz Kwasniewski
 			LUC SBG/TS
 			Universitaire Campus Gebouw D
 			3590 Diepenbeek
 			BELGIUM
 			tel(direct): 011/268315
 			fax	   : 011/268301
 			email      : sergiusz.kwasniewski \\at// luc.ac.be
 		     _________________________________________________