[molpro-user] Convergence problems in MULTI, "WARNING: exit with wrong ci vector!!"

Mon May 22 12:31:02 CEST 2017

Dear all,

I am trying to perform a rather large CASSCF calculations (14 electrons in 14 orbitals) for a transition metal complex, including 10 lowest states in one symmetry and spin: 

{multi
occ,59,51
closed,45,51
wf,206,1,2
state,10
start,2140.2
natorb,2140.2
{iterations;do,uncouple,1,to,40;}
{iterations;dont,internal,1,to,40;}
maxiter,40
}

I have tested many options (see below), but in all cases the iteration procedure in MULTI does not converge and behaves more-or-less like this:

 ITER. MIC  NCI  NEG     ENERGY(VAR)     ENERGY(PROJ)   ENERGY CHANGE     GRAD(0)  GRAD(ORB)   GRAD(CI)     STEP       TIME
   1    5   10    0   -2888.65831267   -2888.65899661   -0.00068394    0.10401421 0.00142009 0.00005114  0.25D-01   2072.98
   2    5   12    0   -2888.65911248   -2888.65914897   -0.00003649    0.01727313 0.00183752 0.00000071  0.75D-02   4633.39
   3    5   13    0   -2888.65916089   -2888.65916485   -0.00000396    0.00551549 0.00172582 0.00000027  0.29D-02   7129.37
 WARNING: exit with wrong ci vector!!
   4    9   13  780   -2888.50043310   -2888.50276409   -0.00233099    0.14909468 0.00149085 0.00000018  0.32D+00   9365.53
   5    4   11    0   -2888.65301376   -2888.65853076   -0.00551701    0.12042313 0.00264229 0.00004756  0.26D+00  11419.07
   6    4   11    0   -2888.65895997   -2888.65913627   -0.00017630    0.02419916 0.00386938 0.00000234  0.47D-01  13733.64
   7    5   12    0   -2888.65916047   -2888.65916548   -0.00000501    0.00570816 0.00087036 0.00000036  0.54D-02  16177.15
 WARNING: exit with wrong ci vector!!
   8    7   12  786   -2888.50045888   -2888.50278318   -0.00232429    0.14825178 0.00435252 0.00000015  0.32D+00  18419.89
   9    4   11    0   -2888.65281628   -2888.65851734   -0.00570106    0.12034652 0.00261341 0.00004805  0.27D+00  20491.25
  10    4   11    0   -2888.65895243   -2888.65913562   -0.00018319    0.02411394 0.00373151 0.00000238  0.48D-01  22809.80
  11    5   11    0   -2888.65916035   -2888.65916545   -0.00000510    0.00568229 0.00087519 0.00000063  0.56D-02  25128.05
 WARNING: exit with wrong ci vector!!
  12    7   12  786   -2888.50047277   -2888.50278714   -0.00231437    0.14806859 0.00359637 0.00000015  0.32D+00  27332.06
  13    4   11    0   -2888.65286565   -2888.65852136   -0.00565571    0.12016013 0.00262332 0.00004798  0.27D+00  29399.38
  14    4   11    0   -2888.65895467   -2888.65913585   -0.00018118    0.02406651 0.00373916 0.00000237  0.48D-01  31700.03
  15    5   12    0   -2888.65916039   -2888.65916546   -0.00000506    0.00567262 0.00088874 0.00000037  0.55D-02  34138.52
 WARNING: exit with wrong ci vector!!
  16    7   12  786   -2888.50048819   -2888.50279081   -0.00230262    0.14783537 0.00352199 0.00000015  0.32D+00  36344.33
  17    4   11    0   -2888.65291533   -2888.65852541   -0.00561008    0.11995809 0.00263353 0.00004788  0.27D+00  38402.12
  18    4   11    0   -2888.65895692   -2888.65913608   -0.00017916    0.02401553 0.00374651 0.00000237  0.48D-01  40697.63
  19    5   12    0   -2888.65916044   -2888.65916547   -0.00000503    0.00566261 0.00090308 0.00000037  0.55D-02  43139.95
 WARNING: exit with wrong ci vector!!
  20    7   12  786   -2888.50050353   -2888.50279445   -0.00229092    0.14760125 0.00345623 0.00000015  0.32D+00  45376.72
  21    4   11    0   -2888.65296457   -2888.65852943   -0.00556486    0.11975525 0.00264392 0.00004779  0.27D+00  47439.36
  22    4   11    0   -2888.65895914   -2888.65913631   -0.00017717    0.02396491 0.00375398 0.00000237  0.47D-01  49754.06

As you can see, every 4 iterations there is a message "WARNING: exit with wrong ci vector!!” and there is jump (the energy change and the gradients suddenly increase by two orders of magnitude). Although the iteration procedure continues, the convergence is extremely slow… and the calculations are not going to converge in a reasonable number of iterations.

Following the manual, I have already tried the following options:
#1. switch off the coupling between CI coefficients and orbital rotations  and disable the simultaneous optimisation of internal orbitals & CI, i.e.: "{iterations;do,uncouple,1,to,40;}" and "{iterations;dont,internal,1,to,40;}"
Without these options it’s even worse: the jump occurs in every single iteration!
#2. temporarily replace CLOSED with FROZEN
#3. change the orbital optimisation method from WERNER to AUGMENT
However, none of these really helped to avoid the oscillations (except for #1, without which the problem was even more severe). 

Is there anything more I can try to converge these CASSCF calculations more smoothly? 
Thank you in advance for your suggestions!

Best regards,
Mariusz Radon

P.S. Perhaps I should add that the active space is chosen consciously and I know these calculations are doable (because I was able to converge them in another program, Molcas; however, I need to perform them also in Molpro because Molcas doesn’t have the NEVPT2 method). Moreover, my initial guess is also fairly reasonably (the natural orbitals for the lowest one of these 10 states, computed with exactly the same active space; so I don’t see how can I improve it further). Also the number of states included (10) is well justified because it accounts for quasi-degeneracy of the excited electronic states in this (qualitatively) octahedral complex. The real symmetry is however much lower, only Ci. The basis set is a mix of cc-pVTZ-DK (on the metal and nearest neighbours) and cc-pVDZ-DK (on the less important atoms).

-- 
Mariusz Radon, Ph.D.
Faculty of Chemistry
Jagiellonian University
ul. Ingardena 3
30-060 Krakow, Poland

phone: 48-12-663-22-31
e-mail: mradon at chemia.uj.edu.pl <mailto:mradon at chemia.uj.edu.pl> 
or mariusz.radon at uj.edu.pl <mailto:mradon at chemia.uj.edu.pl>
web: http://www.chemia.uj.edu.pl/~mradon <http://www.chemia.uj.edu.pl/~mradon>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.molpro.net/pipermail/molpro-user/attachments/20170522/bb581bdc/attachment-0001.html>