[Pw_forum] SCF convergence - proposal

[Konstantin Kudin]

 Hello,

 I've been running ZrO2 oxide unit cells for few days
now, and it appears that the very first convergence is
quite hard to achieve starting from the atomic guess
if the geometry is not too close to the optimized one.
Of course, if I knew the optimized geometry
beforehand, I would not run the whole thing in the
first place!

 So, I used some broadening first, and then imported
that density and potential into an insulator
calculation, and that worked. Perhaps I could also
play with the mixing coefficient, but, again, when
there are several structures to optimize, this is
quite time consuming.

 My proposal is to try to make the convergence more
robust, such that regardless of how things are
started, they always converge. That would make jobs
more "start and forget" kind of proposition.
 
 I've worked on a method that guarantees the SCF
convergence regardless of the starting guess, but I do
not know enough about programming PWSCF to implement
it. {J.Chem.Phys. 116, 8255 (2002).}

 What is needed for the method is the storage of the
several previous densitites and potentials {probably
last 3 ones is fine). Then the mixing coefficients are
determined, and the rho_ = c1*rho_i + c2*rho_i-1 +
c3*rho_i-2. However, there is no need to compute
V(rho_) since it is (roughly) V(rho_) = c1*V(rho_i) +
c2*V(rho_i-1) + c3*V(rho_i-2)
 This V(rho_) is then used to generate bands. The
trick is that c1,c2,c3 are determined from a solid
mathematical approach, and so they always work.

 Right now the code uses fixed mixing (0.7), which is
sort of similar to doing V(rho_) with c1 and c2 always
fixed, and V(c1*rho_i+c2*rho_i-1) is honestly
computed.

 Anybody is interested in trying to get this in the
code?

 Kostya

__________________________________
Do you Yahoo!?
Yahoo! Mail SpamGuard - Read only the mail you want.
http://antispam.yahoo.com/tools