The ABINIT Users Mailing List ( CLOSED )

[Gilles Zerah <gilles.zerah@cea.fr>]

Hello, magnetisor!

Soory for the delay in the reply. There was in fact a bug (found by Xavier)
which was not affecting the final result, but only the self-consistent loop.
In other words, in some cases one would fail to converge, but in case
of convergence the results would be ok.
Attached, you will find a "patch" (for V4.0.3), that is new routines 
correcting the bug
which will be included in the next releases.
The corrected routine is :
Src_common/aprxdr.f
which you can use now.

Also, enclosed is a little routine to play with:
A modified version of Src_4wfs/wfconv.f  (This routine, among other things, 
compute
the initial wave function).
In this modified version, the initial wave function  is rotated 90 degrees 
around z in spin space, that
is x->y . In other words, using this version of the routine,  you should 
obtain exactly the same
calculation
(mutatis mutandis)
with spinat along y than with the original version of the routine but with 
spinat along x.
An example of the log files obtained with a simplified version of your  input 
file (smaller cut-off) is
included.
(patchmagn/test-hgh.log_outJ  and  patchmagn/test-hgh.log_outK) .
You should use this routine only to play, unless you want to look at it, 
which is not difficult: look
at the variable "direction" .

I have also added, for your convenience, the .in and .files files.


Gilles Zerah



> Dear abinitioners,
>
> I have found the following problem.
>
> First let me tell you what I get when I use HGH pseudos: Consider
> vector-magnetization calculations (nspeden=4, nspinor=2) without
> spin-orbit (for simplicity). It turns out that for magnetizations along x
> and z (which we "set" with spinat) abinit works perfectly well. In fact,
> one can nicely reproduce the result obtained with nsppol=2. However, for a
> magnetization along y the calculation does not converge!!
>
> I enclose the files test-hgh.in, test-hgh.out, and test-hgh.log,
> corresponding to rough calculations for MnF2 in ferromagnetic spin
> configurations. The three datasets correspond to magnetizations along the
> three Cartesian directions. The results for x and z reproduce that of the
> nsppol=2 calculation (see enclosed file test-hgh-nsppol2.out). The problem
> for the y direction (dataset # 2, which is calculated in the third place)
> is apparent. The program seems to approach the right solution but it does
> not get to converge and, after 59 steps, it decides to stop
> (``Potential-based CG line minimization not converged after 13
> restarts.'')
>
> ((By the way: The lines in the output and log files that start with ``3D
> magnetization'' report the value of the magnetization vector, and are
> useful to make sure that spinat works. These are written by our local
> version of abinit. Apart from that, the log and output files should be
> reproducible.))
>
> I also enclose results for the same calculations but using
> Troullier-Martins pseudos (files test-tm.out and test-tm-nsppol2.out). The
> x and z cases both converge in 23 steps, while it takes 32 steps for the y
> case to converge. Other tests I have done indicate this is not a
> coincidence.
>
> I have looked into the code but have not been able to find the cause of
> the asymmetry between x and z (which are fine) and y. It occurs to me that
> a magnetization along y is particular in the sense that it requires the
> real part of the first spinor wave function to be related with the
> imaginary part of the second (etcetera), which may make this case be more
> sensitive to certain [coding mistakes? / numerical inaccuracies?] than the
> x and z cases are??
>
> Anyone can help with this?
>
> Thanks,
>
> Jorge
>
> ------------------------------------------------------------------------------
>    Jorge Iniguez             Bldg 235, Room E19
>                              NIST Center for Neutron Research
>                              100 Bureau Drive, Mail Stop 8562
>                              Gaithersburg, MD 20899-8562
>                              Phone: 1-301-975-8367 Fax: 1-301-921-9847
>                              E-mail address: jiniguez@nist.gov
>                             http://www.ncnr.nist.gov/staff/jorge
> ------------------------------------------------------------------------------
>
>   
> ----------------------------------------------------------------------------------------------------
>                    Name: files.tar.gz
>    files.tar.gz    Type: Unix Tape Archive (application/x-tar)
>                Encoding: BASE64

Attachment: patchmagn.tar.gz
Description: GNU Zip compressed data