The ABINIT Users Mailing List ( CLOSED )

["D. R. Hamann" <drh@bell-labs.com>]

Dear Tianshu,

I have two suggestions, but no real idea if they will work.  Try 
increasing tsmear substantially, say to 0.05.  tsmear and 
near-degenerate energy levels at k and k+Q for very small Q may interact 
in funny ways inside the routine vtowfk3.  If this calms the convergence 
issue, try working back down slowly.  Usually increasing the k mesh will 
compensate for making tsmear smaller in GS calculations, but I'm not 
sure that will necessarily be true here.  Stick with occopt=3.

Second suggestion:  Put in another dataset after dataset 1 where you do 
a non-selfconsistent calculation of your GS wave functions on the 
unshifted k grid, using tolwfr=1.0d-20.  The reason for this is that 
higher bands with near-zero occupancy do not effect the self-consistent 
potential, and may not be converged by the tolvrs test in dataset 1.  
This is less likely to be the cause of the convergence problem, although 
it can cause inaccuracies.

Third suggestion, now that I think of it:   Can you converge phonons at 
gamma?  (Be sure you leave frzfermi=0, the default.)  If so, I'd suggest 
you forget near-gamma and use anaddb to interpolate from a uniform 
phonon Q mesh in the irreducible BZ.  If the system is really just one 
atom per unit cell, of course, the only phonon frequency at Q=0 is zero, 
and I'm not sure what Abinit does in this case.  Hopefuly it produces a 
sensible DDB file that can be merged with the other Q's for input to anaddb.

Don Hamann

Tianshu Li wrote:

> Greetings,
>
> I have been suffering from the convergence problem when dealing with 
> the phonon calculation in the transition metal system, say, Vanadium. 
> The symptom I can describe is that the convergence algorithm (CG, 
> Anderson mixing and simple mixing) poorly behaves (it oscillates in CG 
> while totally blows out in Anderson and simple mixing) when the 
> wavevector q is close to the Gamma point. On the other hand, they work 
> fine for the shorter wavelength phonons. I tried something I can think 
> of to improve the convergence behavior, including the different 
> pseudopotential schemes, higher energy cutoff, denser kpoints, more 
> strict convergence tolerance on GS, but they don't seem to help much. 
> (Here comes another issue that the high energy cutoff along with the 
> dense kpoints would dramatically increase the NSC wavefunction storage 
> to 2GB quota that can't be read by the following SC step. How would 
> anybody solve that dilemma?) The other thing I might also try is to 
> raise the number of bands, but given the fact that there are only 5 
> electrons and as many as 14 bands have already been involved, I can 
> hardly see how that would help. From exploring the former archives 
> regarding the similar issues, I know the exchange-correlation and 
> pseudopotential schemes are particularly important for the quality of 
> phonon calculations. But I am not sure how the convergence behavior 
> would rely on the wavevector q. Do I just miss something critical? Can 
> anybody help me out on this issue? Many thanks.
>
> Tianshu Li
> DMSE
> University of California, Berkeley
>
>
>
> Here are the input parameters (Run on MPI Linux clusters)
>
> (LDA+FHI(TM type) pseudopotential)
>
> *INPUT*
> ##########################################################
>    ndtset  3
>
> #Data set 1: Ground state calculation
>   kptopt1   1          # Automatic generation of k points, taking
>                             # into account the symmetry
>     iscf1   5            # CG
>   prtden1   1             # Will be needed for dataset 2
>   tolvrs1   1.0d-20       # SCF stopping criterion
>
>
> #Data set 2: Non-self consistent ground-state calculation
> # with a general q [-0.05 0.05 0.05] along Gamma-X in BCC
>
>     nqpt2   1
>      qpt2   -0.05 0.05 0.05
>   getwfk2   1             # Uses as input wfs the output wfs of the 
> dataset 1
>   getden2   1             # Uses as input density the output density 
> of the dataset 1
>   tolwfr2   1.0d-20       # Stopping tolerance in residual wfn (pretty 
> strict)
>   kptopt2   3             # No symmetry taken into account
>                         
>     iscf2   -2            # Non-self-consistent calculation
>
> #Data set 3: RF calculation
>
>   rfphon3   1             # Activate the calculation of the atomic 
> dispacement perturbations
>  rfatpol3   1 1           # All the atoms will be displaced
>    rfdir3   1 1 1         # x, y, and z
>
>     nqpt3   1
>      qpt3   -0.05 0.05 0.05
>   getwfk3   1             # Uses as input wfs the output wfs of the 
> dataset 1
>   getwfq3   2             # Uses as input ddk wfs the output of the 
> dataset 2
>
>   kptopt3   3           
>   tolvrs3   1.0d-10     # SC
>                        
>     iscf3   5             # CG
>
> # Common input variables
>
> ecut    50.0          
> nstep  70              # Maximal number of SCF cycles
>
> acell 3*5.6020181037 
> rprim   -0.5  0.5  0.5  # BCC
>          0.5 -0.5  0.5
>          0.5  0.5 -0.5
>
>
> kptrlatt  0  8  8      # Tried denser kpoints
>           8  0  8     
>           8  8  0     
>                       
>
>
> ixc 7                 # Perdew-Wang 92 functional, consistent with
>                         # FHI potential (TM scheme)
> occopt 3            # Fermi-Dirac smearing
> tsmear 0.01       # Typical for transition metal
> nband  20          # For phonon calculation, use large number of
>                          # band.
>
> ecutsm  0.5             # Energy cutoff smearing
>
> ntypat 1              
>                     
> znucl  23            # V
>
> natom 1                # There is only one atom V
> typat 1            
> xred                    # This keyword indicate that the location of 
> the atoms
>                            # will follow, one triplet of number for 
> each atom
>    0.0  0.0  0.0       # Triplet giving the REDUCED coordinate of atom 1.


--
D. R. Hamann                            Phone: 908-582-4454
Director, Theoretical Materials         Fax:   908-582-4868
 Physics Research (retired)             email: drh@physics.bell-labs.com
Bell Laboratories 
Lucent Technologies
700 Mountain Ave, Room 1D-371
Murray Hill, NJ 07974-0636 USA