The ABINIT Users Mailing List ( CLOSED )

["Anglade Pierre-Matthieu" <anglade@gmail.com>]

On Dec 2, 2007 5:13 AM, Hua Bao <hbao@purdue.edu> wrote:

Dear Abinit-user,

I am trying to do a phonon dispersion calculation using "respond
function" in Abinit.  I followed the response function 2 in the
tutorial. It used tolvrs 1d-18 for the first dataset, and tolwfr 1d-22
for the second and tolvrs 1d-8 for the others. However, I find that the
convergence criteria is quite difficult to achieve for my system (after
100 steps, tolvrs is only about 1d-2).

I have two questions here.
First, why these criteria are chosen for response function calculation
over toldff and toldfe?

On the one hand, tolvrs is the "real" convergence criterion for SCF loop, because it is the measure of the change of the potential between iterations; because Eigenvalues of the SCF cycles range (almost) between 1 and inifinity, having tolvrs=0 means that your calculation is converged. On the other hand it is quite possible to find not at all converged calculations where toldfe and toldff are zero making them unreliable convergence criterion.

Tolwfr monitor the convergence of wave functions during the non-SCF optimisation. Response function formalism is based on those wave functions (cf some articles of X. Gonze in PRB). If I don't mistake, all RF equations implemented into Abinit become true when the energy  is at its minimum with respect to wavefunctions. Experience tells that  any error in the wavefunctions leads to dramatic errors in the RF quantities (because those are derivatives of the energy).

 

Second, is it OK to use a less strict criteria to achieve a easier
convergence? How much can it affect the output (phonon dispersion)?

2nd question: Dramatically.
1st question: Yet is it OK try to do the calculation with various settings for tolwfr and  monitor  the convergence.
 

My input file is attached.

Good idea: I see that you do not use nnsclo and nline. Try to increase those two variables and you will likely get a much better convergence. For a first guess try nline=15 ...


Regards

PMA