big jumps in E (Re: [Wien] a question on convergence)

[Steven Homolya]

Dear Peter Plaha,

Thank you for your reply.

The charge convergence looks fine and the calculation is always saved between 
separate runs. I have pasted the :DIS and :ITE lines for the first 7 
iterations below.

> > If so I'd be interested in possible causes of big jumps (1000s Ry) in
> > energy between two values. For me this occurs with a 108 atom system,
> > :ENE  : *WARNING** TOTAL ENERGY IN Ry =       -29584.775822
> > :ENE  : *WARNING** TOTAL ENERGY IN Ry =       -29594.018681
> > :ENE  : *WARNING** TOTAL ENERGY IN Ry =       -55274.088758
> > :ENE  : *WARNING** TOTAL ENERGY IN Ry =       -29618.620985
> > :ENE  : *WARNING** TOTAL ENERGY IN Ry =       -29643.144369
> > :ENE  : *WARNING** TOTAL ENERGY IN Ry =       -55273.997039
> > :ENE  : *WARNING** TOTAL ENERGY IN Ry =       -29670.361236
>
> a) How does the charge convergence evolve ? (grep :DIS case.scf)
>    Dos it converge ?
>
Yes:
:DIS  :  CHARGE DISTANCE       0.1332575
:DIS  :  CHARGE DISTANCE       0.1175851
:DIS  :  CHARGE DISTANCE       0.0967242
:DIS  :  CHARGE DISTANCE       0.0856304
:DIS  :  CHARGE DISTANCE       0.0512461
:DIS  :  CHARGE DISTANCE       0.0175887
:DIS  :  CHARGE DISTANCE       0.0086769

After more iterations, the charge distance steadily continues to decrease.

> b) could it be that you have done many iterations without savong the
> results ? grep :ITE case.scf   Do you have the same iteration number twice
> (several times) in the scf file ?
>
Not a problem:
:ITE001:  1. ITERATION
:ITE002:  2. ITERATION
:ITE003:  3. ITERATION
:ITE004:  4. ITERATION
:ITE005:  5. ITERATION
:ITE006:  6. ITERATION
:ITE007:  7. ITERATION

In subsequent cycles the iteration count keeps increasing.

>
> If this does not solve the problem, I need a little more info.
>
I am using the mpi version of w2k with 64 CPUs. The system is 107 Al atoms and 
one Cu, both with RMT's of 2.30, and kgen is told to use 64 k-points. I ran 
the calculation with Rkmax = 8.5 and 9.0. I tried Broyd with 0.05, 0.10, 0.20 
and 0.40. In all cases, the problem remains, though the lower of the two very 
different energies appears to converge to better than 0.0001 Ry.

I have recently encountered the same problem with a 64 atom system (63 Al + 1 
Cu). It is less "jumpy" but after 15 or so iteration the seemingly converged 
(and sensible) energy jumps to a much higher value:

j_28.5_phase_1.scf::ENE  : ********** TOTAL ENERGY IN Ry =       -33904.733165
j_28.5_phase_1.scf::ENE  : ********** TOTAL ENERGY IN Ry =       -33904.733129
j_29.0_phase_1.scf::ENE  : ********** TOTAL ENERGY IN Ry =       -18098.264473
j_29.0_phase_1.scf::ENE  : ********** TOTAL ENERGY IN Ry =       -18103.942640

The jumps occur less frequently for the 64 atom systems. I saw it only a 
couple of times during a series of volume optimisation runs. This was w2k-mpi 
using 36 CPUs, 216 k-points.

Regards,

-- 
Steven Homolya
School of Physics and Materials Engineering
Monash University VIC 3800
Australia
Tel: INT +61 3 9905 3694
Fax: INT +61 3 9905 3637