Page 1 of 1

Non-scf PAW calculation cannot converge.

Posted: Thu Dec 18, 2014 2:10 pm
by sheng
Hi I am using Abinit 7.10.2 and I have problem to converge the non-scf calculation after a successful scf computation. The residm parameter stuck at an order of -2 after a few iterations.

My input files:

Code: Select all

   ndtset  3
   paral_rf2 1
   paral_rf3 1

#Definition of the elementary cell
#*********************************
   rprim 1.0 0.0 0.0
         0.0 1.0 0.0
         0.0 0.0 1.0
   acell 7.6152197110E+00  7.6152197110E+00  7.6152197110E+00    # relaxation value


#Definition of the atoms
#***********************

   xred     0.000000  0.000000  0.000000
            0.500000  0.500000  0.500000
            0.000000  0.500000  0.500000
            0.500000  0.000000  0.500000
            0.500000  0.500000  0.000000

   nband 32   # total valence electron/2 + 10
   natom 5
   ntypat 3
   znucl 56 22 8
   typat 1 2 3 3 3

#Definition of the SCF procedure
#*******************************
   nstep 300
   fftalg 401
   diemac 100000.0

#Definition of the plane wave basis set
#**************************************
   ecut 32
   pawecutdg 90
   ecutsm 0.5
   dilatmx 1.05

   ngkpt 6 6 6
   nshiftk 1
   shiftk 0.5 0.5 0.5

#Ground state calculation
  kptopt1   1             # Automatic generation of k points, taking
                          # into account the symmetry
  toldfe1   1.0d-12       # SCF stopping criterion

#Response Function calculation : d/dk
  rfelfd2   2             # Activate the calculation of the d/dk perturbation
   rfdir2   1 1 1         # Need to consider the perturbation in the x-direction only
                          # This is rather specific, due to the high symmetry of the AlAs crystal
                          # In general, just use rfdir 1 1 1
                          # In the present version of ABINIT (v4.6), symmetry cannot be used
                          # to reduce the number of ddk perturbations
    nqpt2   1
     qpt2   0.0 0.0 0.0   # This is a calculation at the Gamma point
  getwfk2   -1            # Uses as input the output wf of the previous dataset
  kptopt2   2             # Automatic generation of k points,
                          # using only the time-reversal symmetry to decrease
                          # the size of the k point set.
  iscf2  -3
  tolwfr2   1.0d-22       # Must use tolwfr for non-self-consistent calculations
                          # Here, the value of tolwfr is very low.

#Response Function calculation : electric field perturbation and phonons
  rfphon3   1             # Activate the calculation of the atomic dispacement perturbations
 rfatpol3   1 5           # All the atoms will be displaced
  rfelfd3   3             # Activate the calculation of the electric field perturbation
   rfdir3   1 1 1         # All directions are selected. However, symmetries will be used to decrease
                          # the number of perturbations, so only the x electric field is needed
                          # (and this explains why in the second dataset, rfdir was set to 1 0 0).
    nqpt3   1
     qpt3   0.0 0.0 0.0   # This is a calculation at the Gamma point
  getwfk3   -2            # Uses as input wfs the output wfs of the dataset 1
  getddk3   -1            # Uses as input ddk wfs the output of the dataset 2
  kptopt3   2             # Automatic generation of k points,
                          # using only the time-reversal symmetry to decrease
                          # the size of the k point set.
  toldfe3   1.0d-12


# add to conserve old < 6.7.2 behavior for calculating forces at each SCF step
 optforces 1


Thank you.

Re: Non-scf PAW calculation cannot converge.

Posted: Sun Dec 21, 2014 10:28 am
by Jordan
Calculating the ddk in PAW is still a work in progress.
The no convergence of the nscf calculation was found by several people already and is under investigation.

Consider using the berry phase to calculate the born effective charges and dielectric constant.
Then you can add them manually in the DDB file after the RF calculation.

(Try a NC calculation to figure out what to change in the DDB file)

Good luck

Jordan

Re: Non-scf PAW calculation cannot converge.

Posted: Sun Dec 21, 2014 12:10 pm
by sheng
Thank you for your suggestion. However I remember that calculation of ddk by berry phase is not supported in paw, that the berryopt supported is -1, 4, 5....

Am I correct?

Re: Non-scf PAW calculation cannot converge.

Posted: Mon Dec 22, 2014 9:51 am
by Jordan
I don't remember exactely the values for berryopt but for sure you can calculate the Z* and epsilon_inf.
Then you do your dfpt calculation in paw (lda only) and you modify the _DDB file to add the Z* and epsilon_inf
If I remember correlty, the electric field is ipert N+2 so if you have 5 atoms you should add lines with
idir 7 idir 7
idir n idir 7
where idir is 1 2 3 (x y z) and (n in [1 N])

Look in a NC _DDB file to compare what you need to add.

Jorcan

Re: Non-scf PAW calculation cannot converge.

Posted: Mon Dec 22, 2014 5:12 pm
by sheng
Hi do you mean ddb or ddk file? (Your previous post mentioned ddb file) Why is it necessary to add in the extra information? (I don't think I need to use mrgddb for post processing since I need gamma point only)

When you say do dfpt in paw and add in detail to the ddk file (I assume), did you mean using the ddk file obtained using berry phase method? Isn't ddk file is used to calculate born effective charge in the first place?

Thank you in advance.

Re: Non-scf PAW calculation cannot converge.

Posted: Tue Dec 23, 2014 9:08 am
by sheng
Update:
I have tried the same calculation using the same input file with nc pseudopotential but resulting in the same unconverged result. However when I use the berry phase approach the calculation converges and finishes with acceptable results.

I have also tried lda paw as you suggested and found that the calculation can converge and finish with the input file attached. Unfortunately the result obtained differs significantly with those found with nc pseudopotential. The nc results are much closer to experimental results. For example the born effective charge calculated are:

paw nc
Ba: 1.8222 2.7225
Ti: 8.3621 7.4259
O1: -4.3915 -5.9157
O2: -0.8002 -2.1131

Input file with berry approach:

Code: Select all

# Polarization of BaTiO3
# (Sheng, 22.8.2014)
#

   ndtset  3
   symmorphi 0

#Definition of the elementary cell
#*********************************
   rprim 1.0 0.0 0.0
         0.0 1.0 0.0
         0.0 0.0 1.0
   acell 7.6152197110E+00  7.6152197110E+00  7.6152197110E+00    # relaxation value


#Definition of the atoms
#***********************

   xred     0.000000  0.000000  0.000000
            0.500000  0.500000  0.500000
            0.000000  0.500000  0.500000
            0.500000  0.000000  0.500000
            0.500000  0.500000  0.000000

   nband 32   # total valence electron/2 + 10
   natom 5
   ntypat 3
   znucl 56 22 8
   typat 1 2 3 3 3

#Definition of the SCF procedure
#*******************************
   nstep 300
   fftalg 401
   diemac 10.0

#Definition of the plane wave basis set
#**************************************
   ecut 32
   pawecutdg 90
   ecutsm 0.5
   dilatmx 1.05

   ngkpt 6 6 6
   nshiftk 1
   shiftk 0.5 0.5 0.5

#Ground state calculation
  kptopt1   1             # Automatic generation of k points, taking
                          # into account the symmetry
  toldfe1   1.0d-12       # SCF stopping criterion

#DATASET2 : non scf calculation: GS WF in the whole BZ
#*****************************************************
   getden2 1
   getwfk2 1
   kptopt2 2
   iscf2 -2
   tolwfr2 1.0d-22
   berryopt2 -2
   rfdir2 1 1 1

# Response function
#********************
   getwfk3    2
   getddk3    2
   rfphon3    1
  rfatpol3    1 5
   toldfe3    1.0d-12
   kptopt3    2
    rfdir3    1 1 1


# add to conserve old < 6.7.2 behavior for calculating forces at each SCF step
 optforces 1

Re: Non-scf PAW calculation cannot converge.

Posted: Tue Dec 23, 2014 10:19 am
by Jordan
Hi,

Sorry, I meant _DDB file.
After a small converstion to the expected quantity (in the correct unit too) of a _DDB file you should be able to complete the _DDB file knowing the Z* and epsilon.

Results can highly depend on the pseudo potentiel you use and the XC functionnal.
BaTiO3 has widely been studied so you can find the methods and results in litterature.

If you could send your pseudos (NC and PAW or links to the one you used) and input file so we can see if the non convergence of your calculation is bug or not.

Jordan

Re: Non-scf PAW calculation cannot converge.

Posted: Tue Dec 23, 2014 4:12 pm
by sheng
Hi the nc pseudo is the rappe GGA which I got from their website. For the paw pseudo I use J-th LDA pesudopotential (v2) given in Abinit website.

Re: Non-scf PAW calculation cannot converge.

Posted: Sat Jan 10, 2015 8:59 am
by sheng
I found the normal DPFT ddk non-scf iterations can converge if I limit nband to the number of valence bands only.
Then the problem lies in the empty conduction bands.

If I use berry phase to calculate the ddk (with some empty conduction bands), DPFT of phonon perturbation can be done successfully. However, when I combine both of phonon and electric field perturbations, the code will complain that the phonon has zero norm. This error does not appear in I does not include the conduction bands.

So does it mean that in a RF calculation the conduction bands must not be included?