no convergence blues with bulk sio2

[tkuehne]

Hi Axel,

I want to add that OUTER_LOOP, as once described by Joost (email
appended), works well, in particular in conjunction with the FULL_ALL
preconditioner.

MfG,
Thomas

From: "VandeVondele Joost" <von... at pci.unizh.ch>
Date: October 11, 2006 4:57:01 PM GMT+02:00
To: <cp2k... at lists.berlios.de>
Subject: [Cp2k-team] Improved OT (call for testing ..)

Hi All,

I've been working on the preconditioning in OT, and have found very
significant improvements (especially for 'difficult' systems). The
'recommended' (OT related) input section now reads as:

     &QS
       EPS_DEFAULT 1.0E-12
       MAP_CONSISTENT
       EXTRAPOLATION PS
       EXTRAPOLATION_ORDER 2
     &END QS
     &SCF
       SCF_GUESS ATOMIC
       EPS_SCF 1.0E-6
       MAX_SCF 30
       &OUTER_SCF
          EPS_SCF 1.0E-6
          MAX_SCF 20
       &END
       &OT
          MINIMIZER CG
          PRECONDITIONER FULL_ALL
          ENERGY_GAP 0.001
       &END
     &END SCF

I'm thinking about making these (or similar) settings the default,
but
would like some additional testing of them. I've done some tests with
very
impressive results. For example, for an Fe2+ system (i.e. that was
mentioned some time ago on the help list) the number of steps needed
to
converge went from ~1000 to about 80 at the first iteration, while MD
can
is performed with about 7-8 steps. I've seen similar improvements for
some
of my (radical/transition metal) systems. Even 'plain' water does
well,
i.e. MD with just 4/5 SCF steps if DIIS is used (as usual, DIIS is
sometimes faster, but also less robust than CG, bad behaviour seems
rare
with the current 'recommended' settings). I would appreciate some
feedback
on how it works for other systems, so please try it on your favorite
systems and provide some feedback, both positive and negative
feedback
would be appreciated. If you have systems that don't work well, I
would be
happy to get the inputs, so that I can have a look.

There are a few ingredients that make it work :

outer_scf : resets the scf every max_scf and uses the new
wavefunctions to make a more accurate preconditioner
full_all  : now hopefullly numerically more robust, and 'smarter' in
that  it uses and estimate of the error on H_ks to tame the
preconditioner, which allows the use of the 'physical' energy_gap.
energy_gap : now a number that must be smaller than the physical
energy
gap of the system (so metals might still be tricky, the number must
be
nonzero in reality).

The catch is that this preconditioner is 'expensive' to construct
(requires a diagonalization) and somewhat 'expensive' to apply. In my
experience, for most small (<1000 atoms) systems, this is not an issue
as
the gain in number of SCF steps is more important. The things to watch
are
'make_preconditioner' and 'apply_preconditioner' in the timing report.
If
these dominate the computational cost, full_all might not be the
correct
choice for the system.

Let me know,

Joost


-----------------------------------------------------
Dr. Joost VandeVondele,               Office: 34-L-14
Institute of Physical Chemistry, University of Zurich
Winterthurerstrasse 190,     8057 Zurich, Switzerland
Tel. +41 44 6354421             Fax. : +41 44 6356838
-----------------------------------------------------
_______________________________________________
Cp2k-team mailing list
Cp2k... at lists.berlios.de
https://lists.berlios.de/mailman/listinfo/cp2k-team

On Aug 3, 3:06 am, Axel <akoh... at gmail.com> wrote:
> hi all!
>
> the following input fails to converge (either with OT set to ON or
> OFF).
> the exact same setting work very nicely with bulk h2o...
> does anybody have an idea where the problem could be?
>
> &GLOBAL
>   FFTLIB FFTW3
>   PROJECT sio2-liq-qs
>   RUN_TYPE MD
>   PRINT_LEVEL LOW
> &END GLOBAL
>
> &MOTION
>   &MD
>     ENSEMBLE NVE
>     STEPS 1000
>     TIMESTEP 0.2
>     TEMPERATURE 4000.0
>     TEMPERATURE_TOLERANCE 100.0
>   &END MD
>   &PRINT
>     &TRAJECTORY
>     &END TRAJECTORY
>     &VELOCITIES OFF
>     &END VELOCITIES
>   &END PRINT
> &END MOTION
>
> &FORCE_EVAL
>   METHOD QS
>   &DFT
>     CHARGE 0
>     BASIS_SET_FILE_NAME GTH_BASIS_SETS
>     POTENTIAL_FILE_NAME GTH_POTENTIALS
>     &MGRID
>       CUTOFF 300
>     &END MGRID
>     &QS
>       WF_INTERPOLATION ASPC
>       EXTRAPOLATION_ORDER 3
>     &END QS
>     &SCF
>       SCF_GUESS ATOMIC
>       EPS_SCF      1.0E-6
>       MAX_SCF       30
>       WRITE_RESTART_EACH 0
>       &OT OFF
>         MINIMIZER DIIS
>         N_DIIS  7
>       &END OT
>       &OUTER_SCF
>         EPS_SCF 1.0E-6
>         MAX_SCF 5
>       &END
>     &END SCF
>     &XC
>       &XC_FUNCTIONAL PBE
>       &END XC_FUNCTIONAL
>       &XC_GRID
>         XC_SMOOTH_RHO NN50
>         XC_DERIV NN50_SMOOTH
>       &END
>     &END XC
>   &END DFT
>   &SUBSYS
>     &CELL
>       ABC 14.257800 14.257800 14.257800
>       UNIT ANGSTROM
>     &END CELL
>     &TOPOLOGY
>       COORDINATE PDB
>       COORD_FILE_NAME sio2-liq.pdb
>     &END
>     &KIND Si
>       BASIS_SET TZV2P-GTH
>       POTENTIAL GTH-PBE-q4
>     &END KIND
>     &KIND O
>       BASIS_SET TZV2P-GTH
>       POTENTIAL GTH-PBE-q6
>     &END KIND
>   &END SUBSYS
> &END FORCE_EVAL