Re: rasscf maxBLB and maxROT values

Posted by Björn Roos on May 29, 2002 at 07:48:59:

In Reply to: rasscf maxBLB and maxROT values posted by Alessandro Cembran on May 28, 2002 at 10:42:47:

: Dear all,
: I runned a single-state casscf (12,12) on the 1st excited state of a very large system
: (372 basis functions, symmetry=S1).
: The delta E gained quickly the threshold of 1.0E-6, but the maxBLB and maxROT values
: didn't gain the 1.0E-4 threshold. These values didn't decrease below 0.5E-3, even after
: a lot of iteration. I tried to use a level shift of 1.5, but the problem still remained.
: One solution is to raise the thresholds, but this can be dangerous, because I want to run
: the caspt2 on the rasscf wavefunction, and to do so the 1st order wavwfunction should be
: well-converged.
: Any suggestion?

Dear Dr. Cembran,
This can be a tricky problem. The problem with MOLCAS is that it uses only an approximate second order optimization scheme, based on an updated Hessian matrix, which was never computed exactly. Usually the convergence goes to second order at the end of the caclulation, but this problem can sometimes happen.
One thing you could try is to turn off the quasi-newton update of the hessian. This has sometimes helped me in tricky cases. Convergence is normally slower but often steadier,
so I would start from the best converged solution I have (restart with JobIph and Cirestart in the input). The keyword to do this is NOQUNE.

You should also check that your CI is converged properly. Otherwise increase the number of
iterations.You can also use TIGHT in the input to ensure that your CI is properly converged, but this is usually only interesting in the first iterations, where there is no good starting vector for the CI.

If you want you can send me your output and I will take a look at it!
Best Regards
Björn Roos
P.S. Are you in Bologna? If so, we can talk about it between the 10-15 June! I will come to Ciamician one of these days.

Follow Ups:

Post a Followup

Name:
E-Mail:

Subject:

Comments:
: : Dear all, : : I runned a single-state casscf (12,12) on the 1st excited state of a very large system : : (372 basis functions, symmetry=S1). : : The delta E gained quickly the threshold of 1.0E-6, but the maxBLB and maxROT values : : didn't gain the 1.0E-4 threshold. These values didn't decrease below 0.5E-3, even after : : a lot of iteration. I tried to use a level shift of 1.5, but the problem still remained. : : One solution is to raise the thresholds, but this can be dangerous, because I want to run : : the caspt2 on the rasscf wavefunction, and to do so the 1st order wavwfunction should be : : well-converged. : : Any suggestion? : Dear Dr. Cembran, : This can be a tricky problem. The problem with MOLCAS is that it uses only an approximate second order optimization scheme, based on an updated Hessian matrix, which was never computed exactly. Usually the convergence goes to second order at the end of the caclulation, but this problem can sometimes happen. : One thing you could try is to turn off the quasi-newton update of the hessian. This has sometimes helped me in tricky cases. Convergence is normally slower but often steadier, : so I would start from the best converged solution I have (restart with JobIph and Cirestart in the input). The keyword to do this is NOQUNE. : You should also check that your CI is converged properly. Otherwise increase the number of : iterations.You can also use TIGHT in the input to ensure that your CI is properly converged, but this is usually only interesting in the first iterations, where there is no good starting vector for the CI. : If you want you can send me your output and I will take a look at it! : Best Regards : Björn Roos : P.S. Are you in Bologna? If so, we can talk about it between the 10-15 June! I will come to Ciamician one of these days.