[molpro-user] Strange CCSD(T) behavior

Gerald Knizia knizia at theochem.uni-stuttgart.de
Fri Dec 19 21:20:33 GMT 2008

On Friday 19 December 2008 17:39, Steven E. Wheeler wrote:
> I am running into some strange (T) behavior for an open-shell (doublet)
> system.  For certain geometries (for example, displaced geometries in a
> finite difference frequency or optimization calculation) I get a (T)
> correction that is about three times the size of the (T) corrections at
> nearby geometries. [...]
> I'm not convinced this is a Molpro issue (running tests now with ACES),
> but wanted to throw it out on the list to see if anyone had any ideas.
> I have checked for orbital instability issues but have found none.

I personally have never seen such behavior. But if you would provide the full 
input files, I could have a quick look.

Some wild speculations on that:
  o I once saw problems when trying to disable integral compression. Did you 
try the calculation with something like 
(note the non-present compress threshold) instead of the
you gave? Of course this should work, but since tweaking the compression 
accuracy it is a rarely used feature, there is a good chance that occassional 
problems might go by undetected.
About the orbital threshold: We usually use {rhf; accu,18} or similar when 
trying to converge RHF to the limit. But both should work.

  o Sometimes compiler errors occur. In particular, I've seen various versions 
of ifort 10.x generate very wrong code in several occasions. Of course this 
is just a wild guess, but you could try recompiling MOLPRO with a different 
compiler, or compiling tripu.f and uccsd.f with lower optimization settings. 
Did you run "make test"?

  o If running in multiprocessor mode: does it also occur in serial mode?

  o Just to make sure: You did check that the calculated electronic states are 
equal in all cases, or? (I.e., no different states in distorted geometries 
due to changes in molecule symmetry, RHF dipole moments are similar and so 
Gerald Knizia

More information about the Molpro-user mailing list