[Pw_forum] Different convergence behavior in non-collinear SO calculation

[H. Lee]

Hi, everyone:

I have been doing a SO magnetic calculation on large systems (>200 atoms)
using QE v5.1 (official release), and depending on number of pools, I
experienced different convergence behaviors; my calculations have not been
finished up to 100 iterations for both cases.
Although they are small, there are differences in total (absolute)
magnetization and total energy.


For example, I performed the same calculations using (1) 48 cores, 1 pool
and (2) 96 cores, 2 pools.


For (1) 48 cores, 1 pool, the part of results in final SCF step are as
follows:


     total energy              =    -286.41849742 Ry

     Harris-Foulkes estimate   =    -286.41961400 Ry

     estimated scf accuracy    <       0.00021221 Ry


     total magnetization       =     0.02     0.01    -0.01 Bohr mag/cell

     absolute magnetization    =     9.25 Bohr mag/cell


     End of self-consistent calculation


     convergence NOT achieved after 100 iterations: stopping


For (2) 96 cores, 2 pools, the part of results are as follows:


     total energy              =    -286.41898856 Ry

     Harris-Foulkes estimate   =    -286.41919445 Ry

     estimated scf accuracy    <       0.00035870 Ry


     total magnetization       =     0.03    -0.03     0.00 Bohr mag/cell

     absolute magnetization    =     9.10 Bohr mag/cell


     End of self-consistent calculation


     convergence NOT achieved after 100 iterations: stopping


For testing purposes, I also performed a calculation using example 11 with
only modification of "starting_magnetization(1)=1.0" in PW folder of QE
v5.1.

In this case, difference is almost negligible, but there is clearly
difference, e.g., total energy, "avg # of iterations", and "polar coord.: r"
.
For example, I attach the results in iteration 1:

(1) 2 cores and 1 pool:

     Self-consistent Calculation


     iteration #  1     ecut=    45.00 Ry     beta=0.70

     Davidson diagonalization with overlap

     ethr =  1.00E-02,  avg # of iterations =  3.0


 ==============================================================================

     atom number    1 relative position :    0.0000   0.0000   0.0000

     charge :     6.365455

     magnetization :          0.000000   -0.000000    4.270383

     magnetization/charge:    0.000000   -0.000000    0.670868

     polar coord.: r, theta, phi [deg] :     4.270383    0.000000
360.000000


 ==============================================================================


     total cpu time spent up to now is        4.1 secs


     total energy              =    -141.70484542 Ry

     Harris-Foulkes estimate   =    -141.58797205 Ry

     estimated scf accuracy    <       0.30762031 Ry


     total magnetization       =     0.00    -0.00     3.44 Bohr mag/cell

     absolute magnetization    =     3.44 Bohr mag/cell

(2) 4 cores and 2 pools

     Self-consistent Calculation


     iteration #  1     ecut=    45.00 Ry     beta=0.70

     Davidson diagonalization with overlap

     ethr =  1.00E-02,  avg # of iterations =  3.2


 ==============================================================================

     atom number    1 relative position :    0.0000   0.0000   0.0000

     charge :     6.365481

     magnetization :          0.000000   -0.000000    4.270286

     magnetization/charge:    0.000000   -0.000000    0.670850

     polar coord.: r, theta, phi [deg] :     4.270286    0.000000
360.000000


 ==============================================================================


     total cpu time spent up to now is        4.0 secs


     total energy              =    -141.70484225 Ry

     Harris-Foulkes estimate   =    -141.58795749 Ry

     estimated scf accuracy    <       0.30758887 Ry


     total magnetization       =     0.00     0.00     3.44 Bohr mag/cell

     absolute magnetization    =     3.44 Bohr mag/cell


Is this just an inevitable numerical noise?

To my knowledge, it seems that just change of number of pools can't change
the results at each SCF step.

Thanks in advance.

Regards.

H. Lee
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20141009/daf96ba4/attachment.html>