[CPMD-list] Slow convergence with surface calculations

[Ari.P.Seitsonen at iki.fi]

Dear Shiwu,

> I was trying to optimize the wavefunction of a Ag(110) 
> surface with CPMD. The convergence is extemely slow. At some
> stage, the calculation can not gets better, the typical output 
> looks this, 
> ----------------------------------------------------------------------
>  597  3.395E-04   6.826E-06   -3211.139109   -1.347E-05     27.85
>  598  3.350E-04   1.909E-05   -3211.138445    6.640E-04     27.78
>  599  3.353E-04   2.010E-05   -3211.138577   -1.318E-04     26.41
>  ODIIS| Insufficient progress; reset!
>  600  8.653E-04   9.429E-05   -3211.135636    2.941E-03     26.42
>  601  1.010E-02   1.010E-03   -3211.046595    8.904E-02     26.14
>  602  1.548E-03   1.165E-04   -3211.134862   -8.827E-02     26.24
>  603  4.091E-04   3.887E-05   -3211.136805   -1.943E-03     26.94
>  604  4.192E-04   3.997E-05   -3211.136854   -4.923E-05     26.38
>  605  3.237E-04   1.982E-05   -3211.137049   -1.949E-04     26.23
> -----------------------------------------------------------------------
> 
> When running the same cell with FP-LAPW method. The calculation
> converges
> within 15-20 iterations, although each iteration with FLAPW takes
> more than 20 time longer time than an electronic step in CPMD. 
> I understand that slow convergence is not a unfamiliar problem with CPMD
> in a surface calculation. My questions is: Is there any tricks one can 
> play in CPMD to accelerate the convergences? Or we should turn to 
> other methods for a surface calculation? 

You don't use any broadening of the occupation numbers, which is most
likely done in the FLAPW calculation (or maybe you use the tetrahedron
method to integrate over the Brillouin zone). The convergence is VERY
difficult, if not impossible, to achieve without a trick there. Maybe
the mixing of the electron density would work, maybe switching to the
finite temperature formalism (yet this is heavier than pure ODIIS in
the current implementation). However, I would probably first try to
anneal the wave functions in an MD run, and then use the Nosè
thermostat during the MD (in case you are planning to do dynamics on
your system).

    Sincerely,

       apsi

> --------------------------------------------------------------------------
> My input file for a Ag(110) surface is as follows
> 
> # O2 on Ag(110) 5x3x5--12 slab
> &CPMD
>     OPTIMIZE WAVEFUNCTION
>     ODIIS 
>     60
>     STRUCTURE BONDS ANGLES DIHEDRAS 
>     SPLINE QFUNCTIONS POINTS 
>     6000
>     BIG MEMORY  
> &END
> &SYSTEM
>    SYMMETRY
>    8 
>    SURFACE  
>    TESR 
>    4 
>    CELL ABSOLUTE DEGREE 
>    27.299375 23.164300 32.759236 90.000000 90.000000 90.000000           
>    CUTOFF
>    25.
>    POISSON SOLVER MORTENSEN 
> &END
> &ATOMS
> *ag_ps.uspp          BINARY    NEWF 
>   LMAX=D
>    75    
>          0.00000000     0.00000000    16.37961800
>          5.45987500     0.00000000    16.37961800
>         21.83950000     0.00000000    16.37961800
> ...................
>        10.91975000     7.72143326    10.91974522
>        21.83950000    15.44286674    21.83949078
>         5.45987500     7.72143326    10.91974522
> 
> &END
> _______________________________________________

-- 
-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-
 Ari Paavo Seitsonen / Ari.P.Seitsonen at iki.fi / http://www.iki.fi/~apsi/
 Tel +41 1 635 44 97 / Fax +41 1 635 68 38 / GSM +41 79 719 09 35      
 Anschrift: Physikalisch Chemisches Institut (PCI), Universität Zürich (UniZh)
 Indirizzo: Winterthurerstraße 190, CH-8057 Zürich
 Address:   Schweiz / Svizzera / Suisse / Svizra / Switzerland