ABINIT tutorial. Lesson on DFT+DMFT...

A DFT+DMFT calculation for SrVO3.


This lesson aims at showing how to perform a DFT+DMFT calculation using Abinit.

You will not learn here what is DFT+DMFT. But you will learn how to do a DFT+DMFT calculation and what are the main input variables controlling this type of calculation.

It might be useful that you already know how to do PAW calculations using ABINIT but it is not mandatory (you can follow the two lessons on PAW in ABINIT, PAW1 and PAW2). Also the DFT+U lesson in ABINIT might be useful to know some basic variables common to DFT+U and DFT+DMFT.

This lesson should take one hour to complete (especially if you have access to several processors).


Copyright (C) 2000-2017 ABINIT group (BAmadon)

Lesson on DFT+DMFT. Table of content:



1. The DFT+DMFT method: summary and key parameters

The DFT+DMFT method aims at improving the description of strongly correlated systems. Generally, these highly correlated materials contain rare-earth metals or transition metals, which have partially filled d or f bands and thus localized electrons. For further information on this method, please refer to [Georges1996] and [Kotliar2006]. For an introduction to Many Body Physics (Green's function, Self-energy, imaginary time, and Matsubara frequencies), see e.g. [Coleman2015] and [Tremblay2017].

Several parameters (both physical and technical) needs to be discussed for a DFT+DMFT calculation.



Go to the top

2. Electronic Structure of SrVO3 in LDA

You might create a subdirectory of the ~abinit/tests/tutoparal directory, and use it for the tutorial. In what follows, the names of files will be mentioned as if you were in this subdirectory

Copy the files ../Input/tdmft_1.in and ../Input/tdmft_x.files in your Work directory and run ABINIT (as usual, the actual "abinit" command is something like ../../../../src/98_main/abinit):

abinit < tdmft_x.files > log_1

This run should take some time. It is recommended that you use at least 10 processors (and 32 should be fast). It calculates the LDA ground state of SrVO3 and compute the band structure in a second step. The variable pawfatbnd allows to create files with "fatbands" (see description of the variable in the list of variables): the width of the line along each k-point path and for each band is proportional to the contribution of a given atomic orbital on this particular Kohn Sham Wavefunction. A low cutoff and a small number of k-points are used in order to speed up the calculation. During this time you can take a look at the input file. There are two datasets. The first one is a ground state calculations with nnsclo=3 and nline=3 in order to have well diagonalized eigenfunctions even for empty states. In practice, you have however to check that the residue of wavefunctions is small at the end of the calculation. In this calculation, we find 1.E-06, which is large (1.E-10 would be better, so nnsclo and nline should be increased, but it would take more time). When the calculation is finished, you can plot the fatbands for Vanadium and l=2. Several possibilities are available for that purpose. We will work with the simple XMGRACE package (you need to install it, if not already avaialble on your machine).

xmgrace tdmft_1o_DS2_FATBANDS_at0001_V_is1_l0002
The band structure is given in eV.

FatbandV


and the fatbands for Oxygen and l=1 with

 
xmgrace tdmft_1o_DS2_FATBANDS_at0003_O_is1_l0001.

FatbandV


In these plots, you recover the band structure of SrVO3 (see for comparison the band structure of Fig.3 of [Amadon2008]). , and the main character of the bands. Bands 21 to 25 are mainly d and bands 12 to 20 are mainly oxygen p. However, we clearly see an important hybridization. The Fermi level (at 0 eV) is in the middle of bands 21-23.

One can easily check thats bands 21-23 are mainly d-t2g and bands 24-25 are mainly eg: just use pawfatbnd=2 in tdmft_1.in and relaunch the calculations. Then the file tdmft_1o_DS2_FATBANDS_at0001_V_is1_l2_m-2, tdmft_1o_DS2_FATBANDS_at0001_V_is1_l2_m-1 and tdmft_1o_DS2_FATBANDS_at0001_V_is1_l2_m1 give you respectively the xy,yz and xz fatbands (ie d-t2g) and tdmft_1o_DS2_FATBANDS_at0001_V_is1_l2_m+0 and tdmft_1o_DS2_FATBANDS_at0001_V_is1_l2_m+2 give the z2 and z2-y2 fatbands (ie eg).

So in conclusion of this study, the Kohn Sham bands which are mainly t2g are the bands 21,22 and 23.

Of course, it could have been anticipated from classical crystal field theory: the vanadium is in the center of an octahedron of oxygen atoms, so d orbitals are splitted in t2g and eg. As t2g orbitals are not directed toward oxygen atoms, t2g-like bands are lower in energy and filled with one electron, whereas eg-like bands are higher and empty.

In the next section, we will thus use the t2g-like bands to built Wannier functions and do the DFT+DMFT calculation.

Go to the top



3. Electronic Structure of SrVO3: DFT+DMFT calculation

3.1. The input file for DMFT calculation: correlated orbitals, screened Coulomb interaction and frequency mesh

In ABINIT, correlated orbitals are defined using the projected local orbitals Wannier functions as outlined above. The definition requires to define a given energy window from which projected Wannier functions are constructed. We would like in this tutorial, to apply the DMFT method on d orbitals and for simplicity on a subset of d orbitals, namely t2g orbitals (eg orbitals play a minor role because they are empty). But we need to define t2g orbitals. For this, we will use Wannier functions.

As we have seen in the orbitally resolved fatbands, the Kohn Sham wave function contains a important weight of t2g atomic orbitals mainly in t2g-like bands but also in oxygen bands.

So, we can use only the t2g-like bands to define Wannier functions or also both the t2g-like and O-p-like bands.

The first case corresponds to the input file tdmft_2.in. In this case dmftbandi=21 and dmftbandf=23. As we only put the electron interaction on t2g orbitals, we have to use first lpawu=2, but also the keyword dmft_t2g=1 in order to restrict the application of interaction on t2g orbitals.

Notice also that before launching a DMFT calculation, the LDA should be perfectly converged, including the empty states (check nline and nnsclo in the input file). The input file tdmft_2.in thus contains two datasets: the first one is a well converged LDA calculation, and the second is the DFT+DMFT calculation.

Notice the other dmft variables used in the input file and check their meaning in the input variable glossary. In particular, we are using dmft_solv=5 for the dmft dataset in order to use the density-density continuous time quantum monte carlo (CTQMC) solver. (See [Gull2011], as well as the ABINIT 2016 paper [Gonze2016] for details about the CTQMC implementation in ABINIT.)
Note that the number of imaginary frequencies dmft_nwlo has to be set to at least twice the value of dmftqmc_l (the discretization in imaginary time). Here, we choose a temperature of 1200 K. For lower temperature, the number of Matsubara frequencies should be higher.

Here we use a fast calculation, with a small value of the parameters, especially dmft_nwlo, dmftqmc_l and dmftqmc_n.

Let's now discuss the value of the effective Coulomb interaction U (upawu) and J (jpawu). The values of U and J used in ABINIT in DMFT use the same convention as in DFT+U calculations in ABINIT (cf [Amadon2008a]). However, calculations in Ref. [Amadon2008] use for U and J the usual convention for t2g systems as found in [Lechermann2006], Eq. 26 (see also the appendix in [Fresard1997]). It corresponds to the Slater integral F4=0 and we can show that U_abinit=U-4/3 J and J_abinit=7/6 J . So in order to use U=4 eV and J=0.65 eV with these latter conventions (as in [Amadon2008]), we have to use in ABINIT: upawu=3.13333 eV; jpawu=0.75833 eV and f4of2_sla=0.

Now, you can launch the calculation:

Copy the files ../Input/tdmft_2.in and modify tdmft_x.files in your Work directory and run ABINIT:

mpirun -n 32 ../../../tmp/src/98_main/abinit < tdmft_x.files > log_2

3.2. The DFT+DMFT calculation: the log file

We are now going to browse quickly the log file (log_2) for this calculation.

Starting from

     =====  Start of DMFT calculation
	 
we have first the definition of logarithmic grid for frequency, then, after:
  == Prepare data for DMFT calculation
The projection of Kohn Sham wavefunctions and (truncated) atomic orbitals are computed (Eq.(2.1) in [Amadon2012]) and unnormalized orbitals are built (Eq.(2.2) in [Amadon2012]) The occupation matrix in this orbital basis is
  ------ Symetrised Occupation
        0.11142  -0.00000  -0.00000
       -0.00000   0.11142  -0.00000
       -0.00000  -0.00000   0.11142
and the Normalization of this orbital basis is
  ------ Symetrised Norm
        0.65790   0.00000   0.00000
        0.00000   0.65790   0.00000
        0.00000   0.00000   0.65790
Now, let's compare these numbers to other quantities. If the preceding LDA calculation is converged, dmatpuopt=1 is used, and dmftbandi=1 and dmftbandf=nband, then the above Symetrised Occupation should be exactly equal to the occupation matrix given in the usual DFT+U occupation matrix written in the log file (with dmatpuopt=1) (see discussion in [Amadon2012]). In our case, we are not in this case because dmftbandi=21 so this condition is not fulfilled. Concerning the norm if these orbitals, two factors play a role: Next the LDA Green's function is computed.
 =====  LDA Green Function Calculation

Then the Green's function is integrated to compute the occupation matrix. Interestingly, the density matrix here must be equal to the density matrix computed with the unnormalized correlated orbitals. If this is not the case, it means that the frequency grid is not sufficiently large. In our case, we find:

        0.11143   0.00000   0.00000
        0.00000   0.11143   0.00000
        0.00000   0.00000   0.11143
So the error is very small (1.10E-5). As an exercise, you can decrease the number of frequencies and see that the error becomes larger.

Then the true orthonormal Wannier functions are built and the Green's function is computed in this basis just after:

 =====  LDA Green Function Calculation with renormalized psichi
The occupation matrix is now:
        0.16937   0.00000  -0.00000
        0.00000   0.16937   0.00000
       -0.00000   0.00000   0.16937
We see that because of the orthonormalization of the orbitals necessary to built Wannier functions, the occupation matrix logically increases.

Then, after:

 =====  Define Interaction and self-energy
The Interaction kernel is computed from U and J, and the self energy is read from the disk file (if it exists). Then, the Green's function is computed with the self energy and the Fermi level is computed. Then the DMFT Loop starts.
 =====  DMFT Loop starts here   
The log contains a lot of details about the calculation (especially if pawprtvol=3). In order to have a more synthetic overview of the calculation (this is especially useful to detect possible divergence of the calculation), the following command extracts the evolution of the number of electrons (LDA, LDA with Wannier functions, and DMFT number of electrons) as a function of iterations (be careful, all numbers of electron are computed differently as explained in the log file):
grep -e Nb -e ITER log_2

Besides, during each DMFT calculation, there are one or more CTQMC calculations:

Starting QMC  (Thermalization)
For the sake of efficiency, the DMFT Loop is in this calculation done only once before doing again the DFT Loop (cf Fig. 1 of [Amadon2012]). At the end of the calculation, the occupation matrix is written and is:
          -- polarization spin component  1
        0.16843   0.00000  -0.00000
       -0.00000   0.16843  -0.00000
       -0.00000   0.00000   0.16843

We can see that the total number of electron is very close to one and it does not change much as a function of iterations. As an output of the calculation, you can find the self energy in file tdmft_2o_DS2Self-omega_iatom0001_isppol1 and the Green's function is file Gtau.dat.

3.3. The self energy

You can use the self-energy to compute the quasiparticle renormalization weight. We first extract the first six Matsubara frequencies:

head -n 8 tdmft_2o_DS2Self-omega_iatom0001_isppol1 > self.dat
Then we plot the imaginary part of the self-energy (in imaginary frequency):
xmgrace -block self.dat -bxy 1:3
Then using xmgrace, you click on Data, then on Transformations and then on Regression and you can do a 4th order fit as:

self


The slope at zero frequency obtained is 0.82. From this number, the quasiparticle renormalisation weight can be obtained using Z=1/(1+0.82)=0.55.

3.4. The Green's function for correlated orbitals

The impurity (or local) Green's function for correlated orbitals is written in the file Gtau.dat. It is plotted as a function of the imaginary time in the interval [0,β] where β is the inverse temperature (in Hartree). You can plot this Green's function for the six t2g orbitals using e.g xmgrace

xmgrace -nxy Gtau.dat 

Gtau


As the six t2g orbitals are degenerated, the six Green's function must be similar, within the stochastic noise. Moreover, this imaginary time Green's function must be negative and the value of G(β) for the orbital i is equal to the opposite of the number of electrons in the orbital i (-ni). Optionnally, you can check how the Green's function can be a rough way to check for the importance of stochastic noise. For example, change for simplicity the number of steps for the DMFT calculation to 1:
nstep2 1
and then use a much smaller number of steps for the Monte Carlo Solver such as
dmftqmc_n 1.d3
save the previous Gtau.dat file:
cp Gtau.dat Gtau.dat_save
Then relaunch the calculation. After it is completed, compare the new Green's function and the old one with the previous value of dmftqmc_n. Using xmgrace,
xmgrace -nxy Gtau.dat_save -nxy Gtau.dat
one obtains:

Gtau2


One naturally sees that the stochastic noise is much larger in this case. This stochastic noise can induces that the variation of physical quantities (number of electrons, electronic density, energy) as a function of the number of iteration is noisy. Once you have finished this comparison, copy the saved Green's function into Gtau.dat in order to continue the tutorial with a precise Green's function in Gtau.dat:
cp Gtau.dat_save Gtau.dat

3.5. The local spectral function

You can now use the imaginary time Green's function (contained in file Gtau.dat) to compute the spectral function in real frequency. Such analytical continuation can be done on quantum Monte Carlo data using the Maximum Entropy method.

A maximum entropy code has been published recently by D. Bergeron. It can be downloaded here. Please cite the related paper [Bergeron2016] if you use this code in a publication.

The code has a lot of options, and definitely, the method should be understood and the user guide should be read before any real use. It is not the goal of this DFT+DMFT tutorial to introduce to the Maximum Entropy Method (see [Bergeron2016] and references therein). We give here a very quick way to obtain a spectral function. First, you have to install this code and the armadillo library by following the guidelines, and then launch it on the current directory in order to generate the default input file OmegaMaxEnt_input_params.dat.

OmegaMaxEnt
Then edit the file OmegaMaxEnt_input_params.dat, and modify the first seven lines with:
data file: Gtau.dat

OPTIONAL PREPROCESSING TIME PARAMETERS

DATA PARAMETERS
bosonic data (yes/[no]): no
imaginary time data (yes/[no]): yes
Then relaunch the code
OmegaMaxEnt
and plot the spectral function:
xmgrace OmegaMaxEnt_final_result/optimal_spectral_function_*.dat
Change the unit from Hartree to eV, and then, you have the spectral function:

spectralfunction


Even if the calculation is not well converged, you recognize in the spectral functions the quasiparticle peak as well as Hubbard bands at -2 eV and +2.5 eV as in Fig.4 of [Amadon2008].

Go to the top

4. Electronic Structure of SrVO3: Choice of correlated orbitals

Previously, only the t2g-like bands were used in the definition of Wannier functions. If there were no hybridization between t2g orbitals and oxygen p orbitals, the Wannier functions would be pure atomic orbitals and they would not change if the energy window was increased. But there is an important hybridization, as a consequence, we will now built Wannier functions with a large window, by including oxygen p-like bands in the definition of Wannier functions. Create a new input file:

cp tdmft_2.in tdmft_3.in
and use dmftbandi= 12 in tdmft_3.in. Now the code will built Wannier functions with a larger window, including O-p-like bands, and thus much more localized. Launch the calculation after having updated tdmft_x.files (if the calculation is too long, you can decide to restart the second dataset directly from a converged LDA calculation instead of redoing the LDA calculation for each new DMFT calculation).
abinit < tdmft_x.files > log_3
In this case, both the occupation and the norm are larger because more states are taken into account: you have the occupation matrix which is
  ------ Symetrised Occupation
        0.23573  -0.00000  -0.00000
       -0.00000   0.23573  -0.00000
       -0.00000  -0.00000   0.23573
and the norm is:
  ------ Symetrised Norm
        0.78223   0.00000  -0.00000
        0.00000   0.78223  -0.00000
       -0.00000  -0.00000   0.78223
Let us now compare the total number of electron and the norm with the two energy window:
 Energy window:                              t2g-like bands         t2g-like+O-p-like bands
 dmftbandi/dmftbandf:                             21/23                      12/23
 Norm:                                                       0.66                        0.78
 LDA Number of electrons (before ⊥):     0.66(=0.11*6)        1.42(=0.235*6)
 LDA Number of electrons (after  ⊥):       1.02                       1.81
For the large window, as we use more Kohn Sham states, both the occupation and the norm are larger, mainly because of the important weight of d orbitals in the oxygen bands (because of the hybridization). Concerning the norm, remind that in any case, it cannot be larger that 0.86. So as the Norm is 0.78, it means that by selecting bands 12-23 in the calculation, we took into account 0.78/0.86*100=90\% of the weight of the truncated atomic orbital among Kohn Sham bands. Moreover, after orthonormalization, you can check that the difference between LDA numbers of electrons is still large (1.02 versus 1.81), even if the orthonormalization effect is larger on the small windows case. Note that in this particular case, with diagonal matrix, the number of electrons before and after orthonormalization are simply linked by n_before/Norm=n_after, i.e. 1.81≈1.42/0.78 and 1.02≈0.66/0.66

At the end of the DFT+DMFT calculation, the occupation matrix is written and is

          -- polarization spin component  1
        0.29450   0.00000   0.00000
        0.00000   0.29450   0.00000
        0.00000   0.00000   0.29450

Similarly to the previous calculation, the spectral function can be plotted using the Maximum Entropy code: we find a spectral function with an hybridation peak at -5 eV, as described in Fig.5 of [Amadon2008].

spectralfunction


Resolving the lower Hubbard bands would require a more converged calculation.

As above, one can compute the renormalization weight and it gives 0.68. It shows that with the same value of U and J, interactions have a weaker effect for the large window Wannier functions. Indeed, the value of the screened interaction U should be larger because the Wannier functions are more localized (see discussion in [Amadon2008]).

Go to the top



5. Electronic Structure of SrVO3: The internal energy

The internal energy can be obtained with

grep -e ITER -e Internal log_3
and select the second occurrence for each iteration (the double counting expression) which should be accurate with iscf=17 (at convergence both expressions are equals also in DFT+DMFT). So after gathering the data:
  1      -1.51483736718814E+02
  2      -1.51480860837124E+02
  3      -1.51479980721122E+02
  4      -1.51479456233951E+02
  5      -1.51479511038784E+02
  6      -1.51479570943715E+02
  7      -1.51479487485907E+02
  8      -1.51479539558451E+02
  9      -1.51479457525225E+02
 10      -1.51479582334490E+02
You can plot the evolution of the internal energy as a function of the iteration.

internal


You notice that the internal energy (in a DFT+DMFT calculations) does not converge as a function of iterations, because there is a finite statistical noise. So, as a function of iterations, first, the internal energy starts to converge, because the modification of the energy induced by the self-consistency cycle is larger than the statistical noise, but then the internal energy fluctuates around a mean value. So if the statistical noise is larger than the tolerance, the calculation will never converge. So if a given precision on the total energy is expected, a practical solution is to increase the number of Quantum Monte Carlo steps (dmftqmc_n) in order to lower the statistical noise. Also another solution is to do an average over the last values of the internal energy. Note that in version 7.10.5, only the Internal energy has a physical meaning in DFT+DMFT and not Etotal or ETOT.

Go to the top

6. Electronic Structure of SrVO3 in DFT+DMFT: Equilibrium volume

We focus now on the total energy. Create a new input file, tdmft_4.in:

cp tdmft_3.in tdmft_4.in
And use acell=7.1605 instead of 7.2605. Relaunch the calculation and note the Internal energy (grep Internal tdmft_4.out).

Redo another calculation with acell=7.00. Then extract the LDA Internal energy and the DMFT Internal energy (grep Internal tdmft_5.out).

acell   Internal energy LDA    Internal energy DMFT
7.0000   -151.51517               -151.4797
7.1605   -151.52399               -151.4877
7.2605   -151.51515               -151.4795
and then plot DMFT and LDA energies as a function of acell. You will notice that the equilibrium volume is very weakly modified by the strong correlations is this case.

Go to the top

7. Electronic Structure of SrVO3: Conclusion

To sum up, the important physical parameters for DFT+DMFT are the definition of correlated orbitals, the choice of U and J (and double counting). The important technical parameters are the frequency and time grids as well as the number of steps for Monte Carlo, the DMFT loop and the DFT loop.

We showed in this tutorial how to compute spectral functions, quasiparticle renormalization weights and total internal energy.

Go to the top