[QE-users] Difficulty with convergence using tefield

Alex.Durie alex.durie at open.ac.uk
Sun Sep 30 15:02:59 CEST 2018


Dear Lorenzo,


Thank you for your response, yes I got convergence without the field in about 23 iterations, but the estimated error was much better at the earlier iterations than with the field.


I take it from your response you believe the issue lies with the convergence? My k-point grid was defined so that the mesh was equally spaced in reciprocal space, however I understand the system is highly inhomogeneous in the field direction so I will try something like 16x16x16.


I understand what you say about edir, I believe also that the system runs more efficiently with the magnetisation along the z-axis. Do you think overall the code is more efficient with say edir = 3,

nspin = 4, angle1(1) = 90, angle2(1) = 0? Provided that's the correct way to set magnetisation along the x-axis?

Many thanks,

Alex


Subject: Re: [QE-users] Difficulty with convergence using tefield

Message-ID:
        <CAG+GtJcU1YPM9MY5bsn8_p6pdg21FLrLXu9ZWN0vr==jMvho6w at mail.gmail.com>
Content-Type: text/plain; charset="utf-8"

edir=1 and 2 are extremely inefficient in parallel. Use edir=3, rotate your
slab. Does it converge without field? Also, your choic is a of k-points
looks the opposite of correct: you need many points along the field
direction.

--
Lorenzo Paulatto
Written on a virtual keyboard with real fingers

On Sat, 29 Sep 2018, 20:43 Alex.Durie, <alex.durie at open.ac.uk> wrote:

> Dear experts,
>
> I am attempting to perform an scf calculation on a semi-infinite slab of
> ferromagnetic cobalt surrounded by a vacuum with an electric field applied
> perpendicular to the magnetisation, out-of-plane.
>
> I am struggling to get convergence in the scf cycle. I achieved
> satisfactory results using the same geometry and k-points without the field.
>
> I have attempted to run it with ecutwfc = 30, and ecutrho at its default
> value. Now I am attempting ecutwfc = 46, ecutrho = 238 (as recommended in
> the header of the pseudopotential file).
>
> I find in both cases the total energy and Harris-Foulkes estimate are
> usually negative, and the estimated scf accuracy is very large, for each
> cycle.
>
> The runtime for larger ecutwfc and ecutrho, is much longer.
>
> Is there something I'm doing wrong, or do I need to alter my expectations
> of runtime? Please see input file below.
>
> I am running the code with MPI, on 4 cores. At 12 iterations, the CPU time
> is 24737.9 secs, the total energy is -377.46347734 Ry, the Harris-Foulkes
> estimate is -1293.19368730 Ry and the estimated scf accuracy is <
> 3322.28583637 Ry. I have also noticed the "Adding external electric field"
> block is duplicated on each cycle, though I suspect that's normal.
>
> Many thanks
>
> Alex Durie
> PhD student
> The Open University
> United Kingdom
>
> &control
>     calculation='scf'
>     restart_mode='from_scratch',
>     pseudo_dir = '/home/alex/QE/pslibrary.1.0.0/pz/PSEUDOPOTENTIALS/',
>     outdir='./'
>     prefix='co'
>     tefield = .true.
>     dipfield = .true.
> /
> &system
>     ibrav = 8,
>     celldm(1) =4.82388,
>     celldm(2) =8.,
>     celldm(3) =1.,
>     nat= 5,
>     ntyp= 1,
>     ecutwfc = 46.0
>     ecutrho = 238.0
>     nspin = 2
>     occupations='smearing', smearing='cold', degauss=0.02
>     starting_magnetization = -1
>     nbnd = 45
>     edir = 2
>     emaxpos = 0.41
>     eopreg = 0.59
>     eamp = 0.000243
> /
> &electrons
>     electron_maxstep = 500
>     diagonalization='cg'
>     conv_thr = 1.0e-6
>     mixing_beta = 0.3
> /
> ATOMIC_SPECIES
> Co 58.933195 Co.pz-n-kjpaw_psl.1.0.0.UPF
> ATOMIC_POSITIONS {bohr}
> Co   0.0000000      0.0000000      0.000000
> Co   2.4119400      3.4110000      2.411940
> Co   0.0000000      6.8220000      0.000000
> Co   2.4119400     10.2330000      2.411940
> Co   0.0000000     13.6440000      0.000000
> K_POINTS {automatic}
> 16 2 16  0 0 0
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20180930/68026763/attachment.html>


More information about the users mailing list