ABINIT, second lesson of the tutorial:

The H2 molecule, with convergence studies.

This lesson aims at showing how to get converged values for the following physical properties :

the bond length
the atomisation energy

You will learn about the numerical quality of the calculations, then make convergence studies with respect to the number of planewaves and the size of the supercell, and finally consider the effect of the XC functional. The problems related to the use of different pseudopotential are left for another lesson (still to be written ...).

You will also finish to read the abinis_help file.

This lesson should take about 1 hour to be done.

Copyright (C) 2000-2004 ABINIT group (XG,RC)
This file is distributed under the terms of the GNU General Public License, see ~ABINIT/Infos/copyright or http://www.gnu.org/copyleft/gpl.txt .
For the initials of contributors, see ~ABINIT/Infos/contributors .

Help files : New user's guide | Abinis (main) | Abinis (respfn) | Mrgddb | Anaddb

Content of lesson 2

2.0 Summary of the previous lesson
2.1 The convergence in ecut
2.2 The convergence in ecut
2.3 The convergence in acell
2.4 The final calculation in Local (Spin) Density Approximation.
2.5 The use of the Generalized Gradient Approximation.

2.0. Summary of the previous lesson.

We studied the H2 molecule in a big box. We used 10 Ha as cut-off energy, a 10x10x10 Bohr^3 supercell, the local-density approximation (as well as the local-spin-density approximation) in the Teter parametrization (ixc=1, the default), and a pseudopotential from the Goedecker-Hutter-Teter table.

At this stage, we compared our results :

bond length : 1.522 Bohr
atomisation energy at that bond length : 0.1656 Ha = 4.506 eV

with the experimental data (as well as theoretical data using a much more accurate technique than DFT)

bond length : 1.401 Bohr
atomisation energy : 4.747 eV

The bond length is awful (nearly 10% off), and the atomisation energy is a bit too low, 5 % off.

2.1 and 2.2 The convergence in ecut

2.1.a Computing the bond length and corresponding atomisation energy in one run.

Before beginning, you might consider to work in a different subdirectory as for lesson_1. Why not "Work2" ?

Because we will compute many times the bond length and atomisation energy, it is worth to make a single input file that will do all the associated operations. You should try to use 2 datasets (try to combine ~ABINIT/Tutorial/t13.in with ~ABINIT/Tutorial/t15.in !). Do not try to have the same position of the H atom as one of the H2 atoms in the optimized geometry.

The input file ~ABINIT/Tutorial/t21.in is an example of file that will do the job, while ~ABINIT/Tutorial/Refs/t21.out is an example of output file. You might use ~ABINIT/Tutorial/t2x.files as "files" file (do not forget to modify it), although it does not differ from ~ABINIT/Tutorial/t1x.files. The run should take less than one minute.

You should obtain the values :

    etotal1  -1.1058360629E+00
    etotal2  -4.7010531340E-01

and

     xcart1  -7.6091430410E-01  0.0000000000E+00  0.0000000000E+00
              7.6091430410E-01  0.0000000000E+00  0.0000000000E+00

These are similar to those determined in lesson 1, although they have been obtained in one run. You can also check that the residual forces are lower than 5.0d-4. Convergence issues are discussed in section 7 of the abinis_help file.
By the way, you have read all the most important parts of the abinis_help file ! You are missing the sections 2, 5, 8. You are also missing the description of many input variables. We suggest that you finish to read entirely the above-mentioned sections of the abinis_help file now, while the knowledge of the input variables will come in the long run.

2.1.b Many convergence parameters have been already identified. We will focus only on ecut and acell. This is because

the convergence of the SCF cycle and geometry determination are well under control thanks to toldfe, toldff and tolmxf (this might not be the case for other physical properties)
there is no k point convergence study to be done for an isolated system in a big box : no additional information is gained by adding a k-point beyond one
the boxcut value is automatically chosen larger than 2 by ABINIT, see the determination of the input variable "ngfft" by preprocessing
we are using ionmov=3 for the determination of the geometry.

For the check of convergence with respect to ecut, you have the choice between doing different runs of the t21.in file with different values of ecut, or doing a double loop of datasets, as proposed in ~ABINIT/Tutorial/t22.in . The values of ecut have been chosen between 10Ha and 35Ha, by step of 5 Ha. If you want to make a double loop, you might benefit of reading again the double-loop section of the abinis_help file.

2.2.a You have likely seen a big increase of the CPU time needed to do the calculation (now, a few minutes). You should also look at the increase of the memory needed to do the calculation (go back to the beginning of the output file). The output data are as follows :

    etotal11 -1.1058360629E+00
    etotal12 -4.7010531340E-01
    etotal21 -1.1218715957E+00
    etotal22 -4.7529731218E-01
    etotal31 -1.1291943792E+00
    etotal32 -4.7773586216E-01
    etotal41 -1.1326879404E+00
    etotal42 -4.7899907995E-01
    etotal51 -1.1346739190E+00
    etotal52 -4.7972721394E-01
    etotal61 -1.1359660026E+00
    etotal62 -4.8022016187E-01

     xcart11 -7.6091430410E-01  0.0000000000E+00  0.0000000000E+00
              7.6091430410E-01  0.0000000000E+00  0.0000000000E+00
     xcart12  0.0000000000E+00  0.0000000000E+00  0.0000000000E+00
     xcart21 -7.5104996718E-01  0.0000000000E+00  0.0000000000E+00
              7.5104996718E-01  0.0000000000E+00  0.0000000000E+00
     xcart22  0.0000000000E+00  0.0000000000E+00  0.0000000000E+00
     xcart31 -7.3977137323E-01  0.0000000000E+00  0.0000000000E+00
              7.3977137323E-01  0.0000000000E+00  0.0000000000E+00
     xcart32  0.0000000000E+00  0.0000000000E+00  0.0000000000E+00
     xcart41 -7.3304297557E-01  0.0000000000E+00  0.0000000000E+00
              7.3304297557E-01  0.0000000000E+00  0.0000000000E+00
     xcart42  0.0000000000E+00  0.0000000000E+00  0.0000000000E+00
     xcart51 -7.3001593298E-01  0.0000000000E+00  0.0000000000E+00
              7.3001593298E-01  0.0000000000E+00  0.0000000000E+00
     xcart52  0.0000000000E+00  0.0000000000E+00  0.0000000000E+00
     xcart61 -7.2955932741E-01  0.0000000000E+00  0.0000000000E+00
              7.2955932741E-01  0.0000000000E+00  0.0000000000E+00
     xcart62  0.0000000000E+00  0.0000000000E+00  0.0000000000E+00

The corresponding atomisation energies and interatomic distances are :

ecut    atomisation   interatomic distance
(Ha)    energy (Ha)      (Bohr)

10       .1656          1.522
15       .1713          1.502
20       .1737          1.480
25       .1747          1.466
30       .1753          1.460
35       .1756          1.459

In order to obtain 0.2% relative accuracy on the bond length or atomisation energy, one should use a kinetic cut-off energy of 30 Ha. We will keep in mind this value for the final run.

Well, 30 Ha is a large kinetic energy cut-off ! The pseudopotential that we are using for Hydrogen is rather "hard" ...

2.3 The convergence in acell

The same technique as for ecut should be now used for the convergence in acell. We will explore acell starting from 8 8 8 to 18 18 18, by step of 2 2 2. We keep ecut 10 for this study. Indeed, it is a rather general rule that there is little cross-influence between the convergence of ecut and the convergence of acell. The file ~ABINIT/Tutorial/t23.in can be used as an example. The CPU time needed is also in the order of a few minutes. The output data (~ABINIT/Tutorial/Refs/t23.out) are as follows :

    etotal11 -1.1188128742E+00
    etotal12 -4.8074164342E-01
    etotal21 -1.1058360629E+00
    etotal22 -4.7010531340E-01
    etotal31 -1.1039109441E+00
    etotal32 -4.6767804747E-01
    etotal41 -1.1039012761E+00
    etotal42 -4.6743724167E-01
    etotal51 -1.1041439320E+00
    etotal52 -4.6735895144E-01
    etotal61 -1.1042058190E+00
    etotal62 -4.6736729686E-01

     xcart11 -7.8427119905E-01  0.0000000000E+00  0.0000000000E+00
              7.8427119905E-01  0.0000000000E+00  0.0000000000E+00
     xcart12  0.0000000000E+00  0.0000000000E+00  0.0000000000E+00
     xcart21 -7.6091430410E-01  0.0000000000E+00  0.0000000000E+00
              7.6091430410E-01  0.0000000000E+00  0.0000000000E+00
     xcart22  0.0000000000E+00  0.0000000000E+00  0.0000000000E+00
     xcart31 -7.5472620965E-01  0.0000000000E+00  0.0000000000E+00
              7.5472620965E-01  0.0000000000E+00  0.0000000000E+00
     xcart32  0.0000000000E+00  0.0000000000E+00  0.0000000000E+00
     xcart41 -7.5491934758E-01  0.0000000000E+00  0.0000000000E+00
              7.5491934758E-01  0.0000000000E+00  0.0000000000E+00
     xcart42  0.0000000000E+00  0.0000000000E+00  0.0000000000E+00
     xcart51 -7.5427689417E-01  0.0000000000E+00  0.0000000000E+00
              7.5427689417E-01  0.0000000000E+00  0.0000000000E+00
     xcart52  0.0000000000E+00  0.0000000000E+00  0.0000000000E+00
     xcart61 -7.5415539738E-01  0.0000000000E+00  0.0000000000E+00
              7.5415539738E-01  0.0000000000E+00  0.0000000000E+00
     xcart62  0.0000000000E+00  0.0000000000E+00  0.0000000000E+00

The corresponding atomisation energies and interatomic distances are :

acell (Bohr)	atomisation energy (Ha)	interatomic distance (Bohr)
8	.1574	1.568
10	.1656	1.522
12	.1686	1.509
14	.1691	1.510
16	.1694	1.508
18	.1695	1.508

In order to reach 0.2% convergence on the interatomic distance, one needs acell 12 12 12. The atomisation energy needs acell 14 14 14 to be converged at that level. At 12 12 12, the difference is .0009 Ha=0.024eV, which is sufficiently small for practical purposes. We will use acell 12 12 12 for the final run.

For most solids the size of the unit cell will be smaller than that. We are treating a lot of vacuum in this supercell ! So, the H2 study, with this pseudopotential, turns out to be not really easy. Of course, the number of states to be treated is minimal ! This allows to have reasonable CPU time still.

2.4 The final calculation in Local (Spin) Density Approximation.

We now use the correct values of both ecut and acell. Well, you should modify the t23.in file to make a calculation with acell 12 12 12 and ecut 30. You can still use the double loop feature with udtset 1 2 (which reduces to a single loop), to minimize the modifications to the file. The file ~ABINIT/Tutorial/t24.in can be taken as an example of input file, and ~ABINIT/Tutorial/Refs/t24.out as an example of output file.

Since we are doing the calculation at a single (ecut, acell) pair, the total CPU time is not as much as for the previous determinations of optimal values through series calculations. However, the memory needs have still increased a bit.

The output data are :

    etotal11 -1.1329372052E+00
    etotal12 -4.7765320649E-01

     xcart11 -7.2661954446E-01  0.0000000000E+00  0.0000000000E+00
              7.2661954446E-01  0.0000000000E+00  0.0000000000E+00
     xcart12  0.0000000000E+00  0.0000000000E+00  0.0000000000E+00

The corresponding atomisation energy is 0.1776 Ha = 4.833 eV
The interatomic distance is 1.4532 Bohr.
These are our final data for the local (spin) density approximation.

We have used ixc=1 . Other expressions for the local (spin) density approximation

ixc=2, 
3 .. 7

are possible. The values 1, 2, 3 and 7 should give about the same results, since they all start from the XC energy of the homogeneous electron gas, as determined by Quantum Monte Carlo calculations.
Other possibilities

ixc=4, 
5, 6

are older local density functionals, that could not rely on these data.

2.5 The use of the Generalized Gradient Approximation.

We will use the Perdew-Burke-Ernzerhof functional, proposed in Phys. Rev. Lett. 77, 3865 (1996).

In principle, for GGA, one should use another pseudopotential than for LDA. However, for the special case of Hydrogen, and in general pseudopotentials with a very small core (including only the 1s orbital), pseudopotentials issued from the LDA and from the GGA are very similar.

So, we will not change our pseudopotential. This will save us lot of time, as we should not redo an ecut convergence test (ecut is often characteristic of the pseudopotentials that are used in a calculation).

Independently of the pseudopotential, an acell convergence test should not be done again, since the vacuum is treated similarly in LDA or GGA.

So, our final values within GGA will be easily obtained, by setting ixc to 11 in the input file t24.in. See ~ABINIT/Tutorial/t25.in for an example.

    etotal11 -1.1621428502E+00
    etotal12 -4.9869631857E-01

     xcart11 -7.1203906739E-01  0.0000000000E+00  0.0000000000E+00
              7.1203906739E-01  0.0000000000E+00  0.0000000000E+00
     xcart12  0.0000000000E+00  0.0000000000E+00  0.0000000000E+00

The corresponding atomisation energy is 0.1647 Ha = 4.482 eV
The interatomic distance is 1.4241 Bohr.
These are our final data for the generalized gradient approximation.

Once more, here are the experimental data :

bond length : 1.401 Bohr
atomisation energy : 4.747 eV

In GGA, we are within 2% of the experimental bond length, but 5% of the experimental atomisation energy. In LDA, we were within 4% of the experimental bond length, and within 2% of the experimental atomisation energy.

Do not forget that the typical accuracy of LDA and GGA varies with the class of materials studied...