Alex Granovsky
gran@classic.chem.msu.su
John McKelvey asked me to post the following on behalf of him as
he has some strange problems to post on the forum directly:
"An oversimplified comment on GDIIS would be that GDIIS allows a
"free" step because it offers not only an estimate of the next set of
coordinates, along with an estimate of the corresponding gradient.
This pair of estimated coordinates and gradients along with the
previous Hessian allows estimation of the next set of coordinates, for
which an accurate energy and gradients will be calculated, and the
cycle is repeated.. So, even more simplified, for N accurately
computed gradients one gets ~ (2N+1) geometry steps."
John
Kind regards,
Alex Granovsky
On Mon Jun 17 '13 3:02am, Leira Ruth A. F. wrote
------------------------------------------------
>Dear Sirs,
>Thank you for the link and the excerpt! So GDIIS is really effective that it is the default for Firefly 8... Does anyone know when will be the release of the manual?
>Gratefully,
>Leira
>On Sun Jun 16 '13 3:42pm, Thomas wrote
>--------------------------------------
>>In addition to Pedro's paper, below is an excerpt from the upcoming Firefly 8.0.0 manual.
>>
GDIIS (geometry optimization using direct inversion in the iterative sub-space) is an extrapolation method analogous to the DIIS method used for the wavefunction optimization. The implementation used in Firefly is unique and uses ideas presented in references 5 and 6. GDIIS is usually superior to RFO and QA, especially on flat regions on the PES. For this reason, it is the default optimizer. 5. S. Vogel, T. H. Fischer, J. Hutter, H. P. Luthi Int. J. Quantum Chem. 45, 6, 679-688 (1993) 6. H. Sellers Int. J. Quantum Chem. 45, 1, pages 31-41 (1993)