[CPMD-list] linear response convergence

[Daniel Sebastiani]

Hi Silviu,


possibly you are running a system that is particularly difficult to
converge. The preconditioner and the conj-grad implementation (esp. step
length) of the response calculation are a little bit tuned for
well-behaved systems.

As a first attempt, I would recommend to change the CG-ANALYTIC variable
to a larger value, e.g. 999 or so. Then, the step length (for the cg) is
computed each time anew using the quadratic estimation.


> only? Is there a way of changing the linear response solver, i.e.
> choosing among CG/DIIS/ODIIS etc.

I think DIIS is also implemented. The keyword in the &RESP section is...
-- no no, I let you guess what it is. ;-)


> 2. Does the solver related keywords in the the &cpmd section affect in
> any way the linear response solver? My impression was that the answer is
> negative, but I couldn't find any support in the documentation.

your impression is right.

> 3. What would be a safe choice for solver parameters in a
> hard-to-converge linear response case?

a large (integer) value for CG-ANALYTIC and a smaller (real) value for
CG-FACTOR  (0.2 < FACTOR < 1.0) are more conservative values. There are
also some other magic undocumented keywords that change the cg behaviour a
little bit - in which direction is difficult to predict, so these keywords
are not really documented. You can try STATEPREC, for instance.

For a spin=1/2 system, you might actually run better with a restricted
open shell calculation, but this is not implemented in connection with
linear response.

Another thing: your system seems to be isolated, i.e. a molecule in a
supercell. In such a situation I would highly recommend to use the
ISOLATED keyword in the &RESP section. (If SYMMETRY=0, this is
autodetected, but not for any pbc symmetry.)



viel Glueck trotz alledem

Daniel



-- 
---------------------------------------------------------
Daniel Sebastiani              sebastia at mpip-mainz.mpg.de
Max Planck Institute for Polymer Research
Ackermannweg 10                Phone  +49 6131 379 260
55128 Mainz, Germany           Fax    +49 6131 379 100
---------------------------------------------------------