[molpro-user] multi: direct vs. no direct

H. -J. Werner werner at theochem.uni-stuttgart.de
Thu Aug 12 16:11:18 BST 2004

I have tried your inputs on our HP, and both converge to the
same solution. There are some points to notice, however:

1.) You have extremly tight convergence thresholds. In direct
    calculations, this means that also very tight screening thresholds
    must be used, otherwise it might not converge (but in fact it
    does on my machine).

2.) The HF convergence of the direct calculation is affacted by
    the screening. Using tighter thresholds improves convergence
    (e.g., gthresh,thrdscf=1.d-13). This is here important, since
     at the very large distance many integrals are very small.

3.) Your HF solution is ionic and the casscf solution is symmetry
    broken (there is a small dipole moment in y direction). 
    If you compute the triplet state in HF, you avoid the HF
    convergence problems and then of course you get a proper
    neutral state. Furthermore, using this as a starting guess
    in casscf yields a proper solution (not symmetry proken)
    and a lower energy. The reason is that in the first case
    (ionic HF solution), orbitals 8.1 to 11.1 sit on H, and
    since the coupling between H and NH2 is virtually zero,
    the program does not swap these with NH2 orbitals. Thus, all
    these are unoccupied and do not contribute to the energy.
    If one starts with the triplet HF wavefunction, 8.1-10.1
    are ok, but 11.1 is still unoccupied. One can rotate the
    starting orbitals manually, and then an even lower energy
    (optimized -56.18652923) is obtained. The safest way
    to obtain proper orbitals at large distances is either to
    use merge, or first do calculations at shorter distances
    and use these as starting guess.

I cannot help with the problem on your local machine. The
compiler I use is HP F90 v2.6.6.

Best regards
Joachim Werner

On Do, 12 Aug 2004, Fabio Mariotti wrote:

>	Dear Joachim Werner,
>	Here are the two input files I used.
>	I've been testing it in the mean while on a LinuxPC
>	cluster and the two files produced an equivalent
>	output: i.e. convergency.
>	So it looks like it is a machine specific problem.
>	The point will be whether it is due to my
>	local installation or specific to the system.
>	In particular i'm using an HP superdome Itanium2
>	running HP-UX with a locally compiled MOLPRO 2002.6.
>	It can also be a strange case were a small change
>	due to round off at machine precision propagates
>	oddly.
>	I'll check compilation parameters and options
>	and let you know if I have any extra information.
>	(I'll send an email to the user list too if I have
>	relevant infos).
>	Please let me know if you have news about
>	this specific hardware.
>	Thanks and
>	Best Regards
>	Fabio
>H. -J. Werner wrote:
>>This is strange. Could you provide the input please?
>>Joachim Werner
>>On Do, 12 Aug 2004, Fabio Mariotti wrote:
>>>	Dear Molpro Users,
>>>	I'm trying to run some benchmarks on my local HP-UX ia64
>>>	machine to check disk usage etc..
>>>	I'm running at present a CASSCF(MULTI) calculation.
>>>	I'm ruuning the same job (same input) with and without
>>>	GDIRECT keyword.
>>>	In a "normal" run both calculation converge at the same
>>>	value of energy.
>>>	If I add a GTHRESH card, namely:
>>>	gthresh,oneint=1.0E-15,twoint=1.0E-15,zero=1.0E-15,prefac=1.0E-20;
>>>	The GDIRECT calculation fails to converge at the MULTI program
>>>	while the NOT DIRECT converges at the previous energy value.
>>>	How the GTHRESH's thresholds affect the GDIRECT calculation?
>>>	Is that a bug?
>>>	Any idea?
>>>	Thanks
>>>	Fabio
>>>Dr. F.Mariotti - mariotti at ir.phys.chem.ethz.ch
>>>ETH Zurich - Honggerberg, Zurich, Switzerland
>Dr. F.Mariotti - mariotti at ir.phys.chem.ethz.ch
>ETH Zurich - Honggerberg, Zurich, Switzerland

Prof. Hans-Joachim Werner
Institute for Theoretical Chemistry
University of Stuttgart
Pfaffenwaldring 55
D-70569 Stuttgart, Germany
Tel.: (0049) 711 / 685 4400
Fax.: (0049) 711 / 685 4442
e-mail: werner at theochem.uni-stuttgart.de

More information about the Molpro-user mailing list