Problem with geometry optimization

Dear all,

I experienced problems with the geometry optimization of
a simple molecule, allene. They happened for DFT and HF and
various basis sets. So to exclude problems with density
integration and keep things simple, I show results only for
HF/3-21G, limiting the number of optimization cycles to 10.

I did 4 optimizations: two using bohr as input unit, two
angstrom, each one with symmetry turned on or off.

Observations:
1) All optimization seem to make unreasonable displacements
after the first optimization step (with gradients being
fairly small).
2) The problem is much more severe when using bohr as input
unit.
3) Turning off symmetry yields slightly different gradients
in the "bohr" case (by about 10^-7 au), which has fairly
significant influence on the displaced geometry after
optimization cycle 1. This is not seen for the "angstrom"
case.
4) There appears to be an "optimization cycle 0" (at least
I see the sequence SCF -> gradient -> displacement -> SCF ->
gradient -> Output on "optimization cycle 1 -> displacement ->...
For cycle 0, things still seem to be alright.
5) The full point group D2d is recognized only in the "bohr case",
only Cs is recognized in the "Angstrom case". In either case,
only C1 is found after the first displacement.

Note that the displacements for the "bohr case" are really
outrageous, essentially dissociating the molecule. But also
in the "Angstrom case" they are large enough to increase
the energy by about 20 mEh and let the optimization proceed
oscillatorily.

I did the same calculation with Gaussian, which converges
in 4 optimization cycles, lowering the energy by about 1.1 mEh.
Initial energy and gradient norm are both reproduced to better than
10^-7 au by MRCC.

I ran another molecule (benzene), for which I did not spot any
problems.

Any help is greatly appreciated....

Best regards,
Dirk Bakowies

Hi,
could you attach the inputs and outputs?
Best,
Jozsef

Dear Jozsef,

I tried, but somehow it didn't seem to work ....
Trying again...

Dirk

Hi Dirk,

thanks for the report, I've been still working on the fix.
It's really a peculiar one, has something to do with compiler code optimization as well.

Best,
Jozsef

Jozsef,
thanks for the update. And good luck with fixing the issue.
Best,
Dirk

Hi Dirk,
I've found the root cause for the issue you reported, I'll provide the fix soon.

To give you some background: during the back-and-forth conversion between the internal and Cartesian coordinates, the so called Wilson's B-matrix is needed, and to have the generalized inverse of it, the SVD factorization of the matrix is required (there are other ways to have that but this is the most common one). And, there is a threshold for zero eigenvalues, and this needs to be fine-tuned otherwise the geom step will be huge.

Best,
Jozsef