ABINIT, lesson PAW2:

Projector augmented-wave technique : the generation of atomic data files


This lesson aims at showing how to compute atomic data files for the projector-augmented-wave method.

You will learn how to generate the atomic data and what the main variables are to govern their softness and transferability.
It is supposed you already know how to use ABINIT in the PAW case

This lesson should take about 1h30.

Copyright (C) 2005-2013 ABINIT group (MT)
This file is distributed under the terms of the GNU General Public License, see ~abinit/COPYING or http://www.gnu.org/copyleft/gpl.txt .
For the initials of contributors, see ~abinit/doc/developers/contributors.txt .

Contents of lesson PAW2 :


 

1. The PAW atomic dataset - introduction

The PAW method is based on the definition of atomic spheres (augmentation regions) of radius rPAW around the atoms of the system in which a base of atomic partial waves φi, of "pseudized" partial waves~φi, and of projectors~pi (dual to~φi) have to be defined. This set of partial-waves and projectors functions plus some additional atomic data are stored in a so-called PAW dataset. A PAW dataset has to be generated for each atomic species in order to reproduce atomic behavior as accurate as possible while requiring minimal CPU and memory resources in executing ABINIT for the crystal simulations. These two constraints are conflicting.
The PAW dataset generation is the purpose of this tutorial.
It is done according the following procedure (all parameters that define a PAW dataset are in bold):
  1. Choose and define the concerned chemical species (name and atomic number).

  2. Solve the atomic all-electrons problem in a given atomic configuration. The atomic problem is solved within the DFT formalism, using an exchange-correlation functional and either a Schrödinger (default) or scalar-relativistic approximation. It is a spherical problem and it is solved on a radial grid. The atomic problem is solved for a given electronic configuration that can be an ionized/excited one.

  3. Choose a set of electrons that will be considered as frozen around the nucleus (core electrons). The others electrons are valence ones and will be used in the PAW basis. The core density is then deduced from the core electrons wave functions. A smooth core density equal to the core density outside a given rcore matching radius is computed.

  4. Choose the size of the PAW basis (number of partial-waves and projectors). Then choose the partial-waves included in the basis. The later can be atomic eigen-functions related to valence electrons (bound states) and/or additional atomic functions, solution of the wave equation for a given l quantum number at arbitrary reference energies (unbound states).

  5. Generate pseudo partial-waves (smooth partial-waves build with a pseudization scheme and equal to partial-waves outside a given rc matching radius) and associated projector functions. Pseudo partial-waves are solutions of the PAW Hamiltonian deduced from the atomic Hamiltonian by pseudizing the effective potential (a local pseudopotential is built and equal to effective potential outside a rvloc matching radius). Projectors and partial-waves are then orthogonalized with a chosen orthogonalization scheme.

  6. Build a compensation charge density used later in order to retrieve the total charge of the atom. This compensation charge density is located inside the PAW spheres and based on an analytical shape function (which analytic form and localization radius rshape can be chosen).

The user can choose between two PAW dataset generators to produce atomic files directly readable by ABINIT.
The first one is the PAW generator ATOMPAW (originally by N. Holzwarth) and the second one is the Ultra-Soft (US) generator (originally written by D. Vanderbilt). In this tutorial, we concentrate only on ATOMPAW.

It is highly recommended to refer to the following papers to understand correctly the generation of PAW atomic datasets:


[1] "Projector augmented-wave method, P.E. Blochl, Phys. Rev. B 50, 17953 (1994)
[2] "A projector Augmented Wave (PAW) code for electronic structure calculations, Part I : atompaw for generating atom-centered functions", N. Holzwarth et al., Computer Physics Communications, 329 (2001) (might also be available at http://www.wfu.edu/%7Enatalie/papers/pwpaw/atompaw.pdf)
[3] "From ultrasoft pseudopotentials to the projector augmented-wave method", G. Kresse, D. Joubert, Phys. Rev. B 59, 1758 (1999)
[4] "Electronic structure packages: two implementations of the Projector Augmented-Wave (PAW) formalism", M. Torrent et al., Computer Physics Communications 181, 1862 (2010) (might also be available at http://www.wfu.edu/%7Enatalie/papers/PAWform/PAWformman.sdarticle.pdf)
[5] "Notes for revised form of atompaw code", by N. Holzwarth, available at http://www.wfu.edu/%7Enatalie/papers/pwpaw/notes/atompaw/atompawEqns.pdf


 

2. Use of the generation code

Before continuing, you might consider to work in a different subdirectory as for the other lessons. Why not "Work_paw2" ?

Provided that ABINIT has been compiled with the "--with-dft-flavor="...+atompaw"" option, the ATOMPAW code is directly available from command line.
First, just try to type: atompaw
if "atompaw vx.y.z" message appears, everything is fine.
Otherwise, you can try "~abinit_compilation_directory/fallbacks/exports/bin/atompaw-abinit"

In any case, in the following, we name atompaw the ATOMPAW executable.


How to use Atompaw ?
Partial waves φi, PS partial waves~φi and projectors~pi are given in wfn.i files.
Logarithmic derivatives from atomic Hamiltonian and PAW Hamiltonian resolutions are given in logderiv.l files.
A summary of the atomic all-electrons computation and PAW dataset properties can be found in the Atom_name file (Atom_name is the first parameter of the input file).
Resulting PAW dataset is contained in:
Atom_name.XCfunc-paw.abinit file (specific format for ABINIT; present only if requested in inputfile)
Atom_name.atomicdata file (specific format for PWPAW code)
Atom_name.XCfunc.xml file (normalized XML file according to specifications from https://wiki.fysik.dtu.dk/gpaw/setups/pawxml.html).

 

3. First (and basic) PAW dataset for Nickel

Our test case will be NICKEL (1s2 2s2 2p6 3s2 3p6 3d8 4s2 4p0).

In a first stage, copy a simple input file for ATOMPAW in your working directory (find it in ~abinit/doc/tutorial/lesson_paw2/Ni.atompaw.input1). Edit this file.

This file has been built in the following way:

1-All-electrons calculation:

How many electronic states do we need to include in the computation ?
Besides the fully and partially occupied states, it is recommended to add all states that could be reached by electrons in the solid.
Here, for Nickel, the 4p state is concerned. So we decide to add it in the computation.

- A line with the maximum n quantum number for each electronic shell; here "4 4 3" means 4s, 4p, 3d.

- Definition of occupation numbers:
For each partially occupied shell enter the occupation number. An excited configuration may be useful if the PAW dataset is intended for use in a context where the material is charged (such as oxides). Although, in our experience, the results are not highly dependent on the chosen electronic configuration.
We choose here the 3d8 4s2 4p0 configuration. Only 3d and 4p shells are partially occupied ("3 2 8" " and "4 1 0" lines). A "0 0 0" ends the occupation section.
- Selection of core and valence electrons selection: in a first approach, select only electrons from outer shells as valence. But, if particular thermodynamical conditions are to be simulated, it is generally needed to include “semi-core states” in the set of valence electrons. Semi-core states are generally needed with transition metal and rare-earth materials. Note that all wave functions designated as valence electrons will be used in the partial-wave basis.
Core shells are designated by a "c" and valence shells by a "v". All s states first, then p states and finally d states.
Here:
c
c

c

v

c

c

v
v
means:
1s core
2s core

3s core

4s valence

2p core

3p core

4p valence

3d valence

Partial-waves basis generation:

Note that for each angular momentum, valence states already are included in the partial waves basis. Here 4s, 4p and 3d states already are in the basis
For each angular momentum, first add "y" to add an additional partial wave. Then, next line, put the value in Rydberg units. Repeat this for each new partial wave and finally put "n"
In the present file,
y
0.5

n
means that an additional s- partial wave at Eref=0.5 Ry as been added.
y
0.

n
means that an additional p- partial wave at Eref=0. Ry has been added.
y
0.

n
means that an additional d- partial wave at Eref=0. Ry as been added.
Finally, partial waves basis contains two s-, two p-  and two d- partial waves.

At this stage, run atompaw !
For this purpose, simply enter: 
atompaw <Ni.atompaw.input1
Lot of files are produced. We will examine some of them.
A summary of the PAW dataset generation process has been written in a file named Ni (name extracted from first line of input file). Open it. It should look like:
Atom = Ni Z = 28
Perdew - Burke - Ernzerhof GGA Log grid -- n,r0,rmax = 2000 2.2810899E-04 8.0000000E+01
Scalar relativistic calculation -- point nucleus
all-electron results
 core states (zcore) = 18.0000000000000
   1  1  0  2.0000000E+00 -6.0358607E+02
   2  2  0  2.0000000E+00 -7.2163318E+01
   3  3  0  2.0000000E+00 -8.1627107E+00
   5  2  1  6.0000000E+00 -6.2083048E+01
   6  3  1  6.0000000E+00 -5.2469208E+00
valence states (zvale) = 10.0000000000000
   4  4  0  2.0000000E+00 -4.1475541E-01
   7  4  1  0.0000000E+00 -9.0035738E-02
   8  3  2  8.0000000E+00 -6.5223644E-01
evale = -185.182300204924
selfenergy contribution = 8.13253645212050
paw parameters:
    lmax = 2
    rc   = 2.30969849741149
    irc  = 1445
Vloc: Norm-conserving Troullier-Martins form; l= 3;e= 0.0000E+00
Projector method: Bloechl
Sinc^2 compensation charge shape zeroed at rc

Number of basis functions 6
 No.  n   l  Energy          Cp coeff      Occ
  1   4   0  -4.1475541E-01 -9.5091493E+00 2.0000000E+00
  2 999   0   5.0000000E-01  3.2926940E+00 0.0000000E+00
  3   4   1  -9.0035738E-02 -8.9594194E+00 0.0000000E+00
  4 999   1   0.0000000E+00  1.0836820E+01 0.0000000E+00
  5   3   2  -6.5223644E-01  9.1576176E+00 8.0000000E+00
  6 999   2   0.0000000E+00  1.3369075E+01 0.0000000E+00
 evale from matrix elements -1.85182309373359203E+02

The generated PAW dataset (contained in Ni.atomicdataNi.GGA-PBE-paw.abinit or Ni.GGA-PBE.xml file) is a first draft.
Several parameters have to be adjusted, in order to get accurate results and efficient DFT calculations.

Note that only Ni.GGA-PBE-paw.abinit file is directly usable by ABINIT.



 

4. Checking the sensitivity of results to some parameters

Try to select 700 points in the logarithmic grid and check if any noticeable difference in the results appears.
You just have to replace 2000 by 700 in the second line of Ni.atompaw.input1 file. Then run atompaw <Ni.atompaw.input1 again and look at the  Ni file:

evale = -185.182300567432
evale from matrix elements -1.85182301887091256E+02

As you see, results obtained with this new grid are very close to previous ones. We can keep the 700 points grid.

You could decrease the size of the grid; by setting 400 points you should obtain:

evale = -185.182294626845
evale from matrix elements -1.85182337214119599E+02

Small grids give PAW dataset with small size (in kB) and run faster in ABINIT, but accuracy can be affected.

- Note that the final rPAW value ("rc = ..." in Ni file) change with the grid; just because rPAW is adjusted in order to belong exactly to the radial grid. By looking in ATOMPAW user's guide, you can choose to keep it constant.

- Also note that, if the results are difficult to get converged (some error produced by ATOMPAW), you should try a linear grid…


Scalar-relativistic option should give better results than non-relativistic one, but it sometimes produces difficulties for the convergence of the atomic problem (either at the all-electrons resolution step or at the PAW Hamiltonian solution step). If convergence cannot be reached, try a non-relativistic calculation (not recommended for high Z materials)

For the following, note that you always should check the  Ni file, especially the values of valence energy ("evale"). You can find the valence energy computed for the exact atomic problem and the valence energy computed with the PAW parameters ("evale from matrix elements"). These two results should be in close agreement!



 

5. Adjusting partial waves and projectors

Examine the partial-waves, PS partial-waves and projectors.
These are saved in files named wfni, where i ranges over the number of partial waves used, so 6 in the present example. Each file contains 4 columns: the radius in column 1, the partial wave φi in column 2, the PS partial wave~φi in column 3, and the projector~ pi in column 4. Plot the three curves as a function of radius using a plotting tool of your choice.

Here is the first s- partial wave /projector of the Ni example:

First s partial wave

Example: plot the wfn6 file, concerning the second d- partial wave:

2nd d partial wave

This partial wave has been generated at Eref=0 Ry and orthogonalized with the first d- partial wave which has an eigenenergy equal to -0.65Ry (see Ni file). These two energies are too close and orthogonalization process produces "high" partial waves.
Try to replace the reference energy for the additional d- partial wave. For example, put Eref=1. instead of Eref=0. (line 24 of Ni.atompaw.input1 file). Run ATOMPAW again and plot wfn6 file:

2nd d partial wave

Now the PS partial wave and projector have the same order of magnitude !

Note again that you should always check the evale values in Ni file and make sure they are as close as possible.
If not, choices for projectors and/or partial waves certainly are not judicious.



 

6. Examine the logarithmic derivatives

Examine the logarithmic derivatives, i.e., derivatives of an l-state d(log(Ψl(E))/dE computed for the exact atomic problem and with the PAW dataset.
They are printed in the logderiv.l files. Each logderiv.l file corresponds to angular momentum quantum number l, and contains three columns of data: the energy, the logarithmic derivative of the l-state of the exact atomic problem and of the pseudized problem. In our Ni example, l=0, 1 or 2.

The logarithmic derivatives should have the following properties:

Here are the three logarithmic derivative curves for the current dataset:
l=0 log derivativesl=1 log derivativesl=2 log derivatives

As you can see, except for l=2, exact and PAW logarithmic derivatives do not match !
According to the previous remarks, try other values for the references energies of the s- and p- additional partial waves.
First, edit again the Ni.atompaw.input1 file and put Eref=3Ry for the additional s- state (line 18); run ATOMPAW again. Plot the logderiv.0 file. You should get:
l=0 log derivatives

Then put Eref=4Ry for the second p- state (line 21); run ATOMPAW again. Plot again the logderiv.1 file. You should get:
l=1 log derivatives
Now, all PAW logarithmic derivatives match with the exact ones in a reasonable interval.

Note: enlarging energy range of logarithmic derivatives plots

It is possible to change the interval of energies used to plot logarithmic derivatives (default is [-5;5]) and also to compute them at more points (default is 200). Just add the following keywords at the end of the SECOND LINE of the input file:
logderivrange -10 10 500
In the above example ATOMPAW plots logarithmic derivatives for energies in [-10;10] at 500 points.

Additional information concerning logarithmic derivatives:
Another possible problem could be the presence of a discontinuity in the PAW logarithmic derivative curve at an energy where the exact logarithmic derivative is continuous.
This generally shows the presence of a “ghost state”.
* Change the pseudopotential scheme. Norm-conserving pseudopotentials are sometimes so deep (attractive near r=0) that they produce ghost states.
A first solution is to change the quantum number used to generate the norm-conserving pseudopotential. But this is generally not sufficient.
A second solution is to select an "ultrasoft" pseudopotential, freeing the norm conservation constraint (simply replace "troulliermartins" by "ultrasoft" in input file)
A third solution is to select a simple “bessel” pseudopotential (replace "troulliermartins" by "bessel" in input file). But, in that case, one has to noticeably decrease the matching radius rVloc if one wants to keep reasonable physical results. Selecting a value of rVloc between 0.6*rPAW and 0.8*rPAW is a good choice; but the best way to adjust rVloc value is to have a look at the two values of evale in Ni file which are sensitive to the choice of rVloc. To change the value of  rVloc, one has to detail the line containing all radii (rPAW, rshape, rVloc and rcore); see user's guide.
* Change the matching radius rc for one (or both) l partial-wave(s). In some cases, changing rc can remove ghost states.
To see an example of ghost state, use the ~abinit/doc/tutorial/lesson_paw2/Ni.ghost.atompaw.input file and run it with ATOMPAW.
Look at the l=1 logarithmic derivatives (logderiv.1 file). They look like:
Ni l=1 log derivatives
Now, edit the Ni.ghost.atompaw.input file and replace "troulliermartins" by "ultrasoft". Run ATOMPAW again... and look at logderiv.1 file. The ghost state has moved !
Edit again the file and replace 'ultrasoft" by "bessel"; then change the 17th line ("2.0 2.0 2.0 2.0") by "2.0 2.0 1.8 2.0". This has the effect of decreasing the rVloc radius. Run ATOMPAW: the ghost state disappears !
Start from the original state of Ni.ghost.atompaw.input file and put 1.8 for the matching radius of p- states (put 1.8 on lines 31 and 32). Run ATOMPAW: the ghost state disappears !


 

7. Testing efficiency of PAW dataset

Let's use again our Ni.atompaw.input1 file for Nickel (with all our modifications).
You get a file Ni.GGA-PBE-paw.abinit containing the PAW dataset designated for ABINIT.

Now, one has to test the efficiency of the generated PAW dataset. We finally will use ABINIT !
You are about to run a DFT computation and determine the size of the plane wave basis needed to get a given accuracy. If the cut-off energy defining the plane waves basis is too high (higher than 20 Hartree, if rPAW has a reasonable value), some changes have to be made in the input file.

Copy ~abinit/tests/tutorial/Input/tpaw2_x.files and ~abinit/tests/tutorial/Input/tpaw2_1.in in your working directory. Edit ~abinit/tests/tutorial/Input/tpaw2_1.in, and activate the eight datasets (only one is kept by default for testing purposes). Run ABINIT with them.
ABINIT computes the total energy of ferromagnetic FCC Nickel for several values of ecut.
At the end of output file, you get this:

      ecut1   8.00000000E+00 Hartree
      ecut2   1.00000000E+01 Hartree
      ecut3   1.20000000E+01 Hartree
      ecut4   1.40000000E+01 Hartree
      ecut5   1.60000000E+01 Hartree
      ecut6   1.80000000E+01 Hartree
      ecut7   2.00000000E+01 Hartree

      ecut8   2.20000000E+01 Hartree
    etotal1  -3.9300291581E+01
    etotal2  -3.9503638785E+01
    etotal3  -3.9583278145E+01
    etotal4  -3.9613946329E+01
    etotal5  -3.9623543087E+01
    etotal6  -3.9626889070E+01
    etotal7  -3.9628094989E+01
    etotal8  -3.9628458879E+01


etotal convergence (at 1 mHartree) is achieve for 18<=ecut<=20 Hartree
etotal
convergence (at 0,1 mHartree) is achieve for ecut>22 Hartree

This is not a good result for a PAW dataset; let's try to optimize it.
Vanderbilt projectors generally are more localized in reciprocal space than Bloechl ones (see ref. [4] for a detailed description of Vanderbilt projectors).
Keyword "bloechl" has to be replaced by "vanderbilt" in the ATOMPAW input file and rc values have to be added at the end of the file (one for each PS partial wave).
You can have a look at the ATOMPAW input file :  ~abinit/doc/tutorial/lesson_paw2/Ni.atompaw.input.vanderbilt
But we will not test this case here as it produces problematic results for this example (see below).
Use this input file for ATOMPAW: ~abinit/doc/tutorial/lesson_paw2/Ni.atompaw.input2
As you can see (by editing the file)  “bloechl” has been changed by ”custom rrkj” and 6 rc values have been added at the end of the file; each one correspond to the matching radius of one PS partial wave.
Repeat the entire procedure (ATOMPAW + ABINIT)... and get a new ABINIT output file.
Note: you have to look again at log derivatives in order to verify that they still are correct...

      ecut1   8.00000000E+00 Hartree
      ecut2   1.00000000E+01 Hartree
      ecut3   1.20000000E+01 Hartree
      ecut4   1.40000000E+01 Hartree
      ecut5   1.60000000E+01 Hartree
      ecut6   1.80000000E+01 Hartree
      ecut7   2.00000000E+01 Hartree
      ecut8   2.20000000E+01 Hartree
    etotal1  -3.9600401638E+01
    etotal2  -3.9627563690E+01
    etotal3  -3.9627901781E+01
    etotal4  -3.9628482371E+01
    etotal5  -3.9628946655E+01
    etotal6  -3.9629072497E+01
    etotal7  -3.9629079826E+01
    etotal8  -3.9629097793E+01
etotal convergence (at 1 mHartree) is achieve for 12<=ecut<=14 Hartree
etotal
convergence (at 0,1 mHartree) is achieve for 16<=ecut<=18 Hartree
This is a reasonable result for a PAW dataset !
Edit ~abinit/doc/tutorial/lesson_paw2/Ni.atompaw.input2 and replace "custom rrkj" by "custom polynom2 7 10"
Repeat the entire procedure (ATOMPAW + ABINIT)... and look at ecut convergence...
    


Optional exercise: let's go back to Vanderbilt projectors

Repeat the procedure (ATOMPAW + ABINIT) with ~abinit/doc/tutorial/lesson_paw2/Ni.atompaw.input.vanderbilt file.
As you can see ABINIT convergence cannot be achieved !
You can try whatever you want with radii and/or references energies in the ATOMPAW input file: ABINIT always diverges !
The solution here is to change the pseudization scheme for the local pseudopotential.
Try to replace the "troulliermartins" keyword by "ultrasoft". Repeat the procedure (ATOMPAW + ABINIT).
ABINIT can now reach  convergence !
Results are below:

      ecut1   8.00000000E+00 Hartree
      ecut2   1.00000000E+01 Hartree
      ecut3   1.20000000E+01 Hartree
      ecut4   1.40000000E+01 Hartree
      ecut5   1.60000000E+01 Hartree
      ecut6   1.80000000E+01 Hartree
      ecut7   2.00000000E+01 Hartree
      ecut8   2.20000000E+01 Hartree
    etotal1  -3.9609714395E+01
    etotal2  -3.9615187859E+01
    etotal3  -3.9618367959E+01
    etotal4  -3.9622476129E+01
    etotal5  -3.9624707476E+01
    etotal6  -3.9625234480E+01
    etotal7  -3.9625282524E+01
    etotal8  -3.9625330757E+01
etotal convergence (at 1 mHartree) is achieve for 14<=ecut<=16 Hartree
etotal convergence (at 0,1 mHartree) is achieve for 20<=ecut<=22 Hartree
Note: You could have tried the "bessel" keyword instead of "ultrasoft"...

Summary of convergency results:

ecut convergence

Final remarks:

 

8. Testing against physical quantities

Finally, the last step is to examine carefully the physical quantities obtained with the PAW dataset.

Copy ~abinit/tests/tutorial/Input/tpaw2_2.in in your working directory. Edit it, to activate the eight datasets (instead of one).
Use the 
~abinit/doc/tutorial/lesson_paw2/Ni.GGA-PBE-paw.abinit.rrkj psp file (it has been obtained from Ni.atompaw.input2 file).
Modify tpaw2_x.files file according to these new files.
Run ABINIT (this may take a while...).

ABINIT computes the converged ground state of ferromagnetic FCC Nickel for several volumes around equilibrium.

Plot the etotal vs acell curve:
etotal vs acell
From this graph and output file, you can extract some physical quantities:
Equilibrium cell parameter: a0 = 3.523 angstrom
Bulk modulus: B = 190 GPa
Magnetic moment at equilibrium: μ = 0.60

Compare these results with published results:
a0 = 3.52 angstrom
B = 200 GPa
μ = 0.60
a0 = 3.52 angstrom
B = 194 GPa
μ = 0.61
a0 = 3.52 angstrom
B = 183 GPa

You should always compare results with all-electrons ones (or other PAW computations), not with experimental ones...


Additional remark

It can be useful to test the sensitivity of results to some ATOMPAW input parameters (see user's guide for details on keywords):

All these parameters have to be meticulously checked, especially if the PAW dataset is used for non-standard solid structures or thermodynamical domains.

Optional exercise: let's add 3s and 3p semi-core states in PAW dataset !

Repeat the procedure (ATOMPAW + ABINIT) with ~abinit/doc/tutorial/lesson_paw2/Ni.atompaw.input.semicore file...
The run is a bit longer as more electrons have to be treated by ABINIT.

Look at
a0, B or μ variation.
Note: this new PAW dataset has a smaller rPAW radius (because semi-core states are localized).
a0 = 3.519 angstrom
B = 194 GPa
μ = 0.60


 

8. The Real Space Optimization (RSO) - experienced users

In this section, an additional optimization of the atomic data is proposed which can contribute, in some cases, to an acceleration of the convergence on ecut. This optimization is not essential to produce efficient PAW datasets but it can be useful. We advise experienced users to try it.

The idea is quite simple: when expressing the different atomic radial functions (φi,~φi~pi) on the plane waves basis, the number of plane waves depends on the "locality" of these radial functions in reciprocal space.
In the following reference (we suggest to read it):  R.D. King-Smith, M.C. Payne, J.S. Lin, Phys. Rev. B 44, 13063 (1991)

A method to enforce the locality (in reciprocal space) of projectors~pi is presented:
Projectors~pi(g) expressed in reciprocal space are modified according to the following scheme:
The reciprocal space is divided in 3 regions:
- If  g < gmax,~pi(g) is unchanged
- If  g > γ,~pi(g) is set to zero
- If gmax< g < γ,~pi(g) is modified so that the contribution of~pi(r) is conserved with an error W (as small as possible).
RSO
The above transformation of~pi(g) is only possible if~pi(r) is defined outside the augmentation sphere up to a radius R0 (with R0>rc).
In practice we have to:
- Impose an error W (W is the maximum error admitted on total energy)
- Adjust gmax according to Ecut (gmax<= Ecut)
- Choose γ so that 2*gmax < γ < 3*gmax
and the ATOMPAW code apply the transformation to~pi and deduce R0 radius.

You can test it now.
In your working directory, re-use the dataset with Bloechl projectors ( ~abinit/doc/tutorial/lesson_paw2/Ni.atompaw.input3).
Replace the last line but one ("default") by "rsoptim 8. 2 0.0001" (8., 2 and 0.0001 are the values for gmaxγ/gmax and W).
Run ATOMPAW.
You get a new psp file for ABINIT.
Run ABINIT with it using the ~abinit/tests/tutorial/Input/tpaw2_1.in file.
Compare the results with those obtained in section 7.

You can try several values for gmax (keeping γ/gmax and W constant) and compare the efficiency of the atomic data; do not forget to test physical properties again.

RSO comparison

How to choose the RSO parameters ?

γ/gmax=2 and 0.0001 < W < 0.001 is a good choice.
gmax has to be adjusted. The lower gmax the faster the convergence is ; but too low gmax can produce unphysical results.