PHYSICAL REVIEW B 84, 085145 (2011)

First-principles study of the band structure and optical absorption of CuGaS2

Irene Aguilera,1,2,3 Julien Vidal,2,3,4 Perla Wahnón,1 Lucia Reining,2,3 and Silvana Botti2,3,5,*
1
Instituto de Energı́a Solar and Departamento de Tecnologı́as Especiales, ETSI Telecomunicación, UPM, Ciudad Universitaria,
ES-28040 Madrid, Spain
2
Laboratoire des Solides Irradiés (LSI), École Polytechnique, CNRS, CEA-DSM, F-91128 Palaiseau, France
3
European Theoretical Spectroscopy Facility (ETSF)
4
Institute for Research and Development of Photovoltaic Energy (IRDEP), UMR 7174 CNRS/EDF/ENSCP,
6 Quai Watier, F-78401 Chatou, France
5
Laboratoire de Physique de la Matière Condensée et Nanostructures (LPMCN), Université “Claude Bernard” Lyon 1, CNRS,
Domaine Scientifique de la Doua, F-69622 Villeurbanne, France
(Received 13 January 2011; revised manuscript received 19 April 2011; published 30 August 2011)
CuGaS2 is the most promising chalcopyrite host for intermediate-band thin-film solar cells. Standard Kohn-
Sham density functional theory fails in describing the band structure of chalcopyrite materials, due to the strong
underestimation of the band gap and the poor description of p-d hybridization, which makes it inadvisable to use
this approach to study the states in the gap induced by doping. We used a state-of-the-art restricted self-consistent
GW approach to determine the electronic states of CuGaS2 : in the energy range of interest for optical absorption,
the GW corrections shift the Kohn-Sham bands almost rigidly, as we proved through analysis of the effective
masses, bandwidths, and relative position of the conduction energy valleys. Furthermore, starting from the GW
quasiparticle bands, we calculated optical absorption spectra using different approximations. We show that the
time-dependent density functional theory can be an efficient alternative to the solution of the Bethe-Salpeter
equation when the exchange-correlation kernels derived from the Bethe-Salpeter equation are employed. This
conclusion is important for further studies of optical properties of supercells including dopants.

DOI: 10.1103/PhysRevB.84.085145 PACS number(s): 71.10.−w, 71.15.Mb, 71.15.Qe, 78.20.Ci

I. INTRODUCTION from photoreflectance and absorption spectra. One should keep

in mind when comparing calculated quasiparticle (QP) energy
In the quest of lower-cost solar cells, the thin-film approach
bands and optical data that the QP (or photoemission) and
has been proposed to cut the costs associated with the
optical gaps differ by the exciton binding energy (<0.05 eV
absorber by reducing the volume of material required. In this
for CuGaS2 ).11
context, Cu-based chalcopyrite semiconductors of the form
According to the Shockley-Queisser limit,13 the maximum
Cu(Ga,In)(S,Se)2 (CIGS) appear to be one of the most promis-
efficiency for solar cells based on p-n junctions can be obtained
ing thin-film alternatives to silicon.1 Compounds belonging
for a band gap of 1.4 eV. CuGaS2 should therefore be ruled
to the CIGS family exhibit structural, optical, and electrical
out as an absorber for single-junction cells.14 On the other
properties that can be, in principle, tuned as a function of
hand, thanks to its wide gap, CuGaS2 is currently used as a
the composition (presence of native defects) and the growth
process. In addition, this class of semiconductors possesses di- window layer for photovoltaic applications. However, the main
rect band gaps and high optical absorption coefficients, which interest of this compound lies in the possibility of hosting a
allows us to reduce to a few microns the thickness necessary narrow intermediate band inside the band gap, when it is highly
to capture all the incoming photons with a frequency higher doped with transition metals (such as Ti or Cr).15–17 In 1997,
than the band gap. In addition to the decrease in the cost of the it was shown18 that a significant increase in the efficiency of
absorber material, this improves the efficiency of the devices, solar cells could be obtained by introducing impurity energy
as higher absorption leads to a higher photocurrent. Thin-film levels in the semiconductor band gap that absorbs additional
chalcopyrite technology is thus nowadays a solid candidate for lower energy photons. It was proposed, and recently observed
second-generation photovoltaics, with the highest efficiency experimentally,19 that when the concentration of impurities is
among thin-film solar cells (20.1%).2 high enough to give rise to a band (the intermediate band),
All chalcopyrite CIGS unit cells present an internal dis- nonradiative recombination can be suppressed, leading to the-
tortion with respect to the standard zincblende lattice, since oretical efficiencies much higher than those of a conventional
the existence of two different cations (Cu, In) results in two solar cell. Since the theoretical optimal value of the band
different bonding lengths between cations and anions.3 The gap for a thin-film host semiconductor for intermediate-band
anion displacement parameter u measures this distortion of photovoltaic applications is around 2.4 eV,20 CuGaS2 is the
the (S,Se) sublattice. It is well known that the value of u has a ideal chalcopyrite host.14–17 Moreover, CuGaS2 was the first
large influence on the band gap of CIGS chalcopyrites.3–6 semiconductor proposed to combine the intermediate-band
The experimental band gaps of CIGS range from 1.05 eV concept and the thin-film technology.
for CuInSe2 7 to around 2.5 eV for CuGaS2 .8 This work is Regarding the optical properties of CuGaS2 , some
devoted to the study of the latter, which has an experimental experimental7,10,11,21,22 and theoretical23–25 studies in the
optical gap of 2.4–2.5 eV8 at room temperature, obtained from visible range have been reported so far. The excitonic
electroreflectance spectra, and 2.53 eV at 0 K,9–12 obtained peak in the optical spectra has been characterized only

1098-0121/2011/84(8)/085145(9) 085145-1 ©2011 American Physical Society

AGUILERA, VIDAL, WAHNÓN, REINING, AND BOTTI PHYSICAL REVIEW B 84, 085145 (2011)

experimentally.10,21,26,27 Moreover, experimental results are In the present work, we use a simplified scGW scheme
not always easy to interpret, in particular, as there are large based on Hedin’s Coulomb hole and screened exchange (COH-
discrepancies among them. From the theoretical side, and SEX) approximation,31 followed by a perturbative G0 W0 step.
despite many efforts, calculations in the literature are not The COHSEX is a static and Hermitian approximation to
accurate enough to help in the understanding of experimental the GW self-energy. scCOHSEX leads to band gaps that are
measurements. Indeed, the main problem comes from the fact usually too large. The final step of our method, the perturbative
that the CuGaS2 electronic states close to the energy gap arise G0 W0 performed on top of the scCOHSEX iterations, accounts
from the hybridization of the d states of Cu with the p states for the dynamical screening that is missing in the COHSEX
of S. This leads to subtle exchange and correlation effects that approximation, yielding QP energies in excellent agreement
are very hard to take into account using standard density func- with experiments.6,35,40 Moreover, scCOHSEX wave functions
tionals. As a result, the p-d hybridization is overestimated.28 were found35 to be very similar to those obtained by applying
The ab initio results published up to now were obtained the original QP scGW scheme (QPscGW ) of Faleev et al.43
from density functional theory (DFT),29,30 based either on the Besides being in good agreement with both experiments
local density approximation (LDA) or on generalized gradient and QPscGW results, the scCOHSEX + G0 W0 scheme is
approximations (GGAs). However, DFT is hampered by two computationally less expensive, as it involves only sums over
important shortcomings when applied in this context: (i) the occupied states to build the self-energy matrix elements.
Kohn-Sham (KS) band gap is systematically underestimated Standard KS DFT is inadequate for the accurate prediction
by 50% to 100% compared to photoemission experiments; of electronic excitations. To calculate optical spectra one has to
and (ii) there is deficient cancellation of the spurious self- deal with neutral excitations due to the simultaneous creation
interaction terms in standard functionals, particularly critical of a quasielectron and a quasihole which interact in the system.
for d electrons, which are usually located too high in energy. A possible way to go beyond the application of Fermi’s
For systems with shallow d states, like the chalcopyrites, this golden rule, which is a simple sum over independent transi-
has a direct effect on the band gap. tions between KS states, is to use the time-dependent (TD)
It is clear that, in order to predict correctly both band gaps generalization of DFT (TDDFT),44,45 where the electronic
and optical absorption spectra, it is necessary to go beyond density responds to a TD external field. TDDFT retains all the
standard KS DFT methods. Many-body perturbation theory advantages of the DFT formalism in terms of computational
(MBPT), in particular, the GW approximation31 for QP states efficiency. However, the lack of good approximations for the
and the subsequent solution of the Bethe-Salpeter equation TD exchange-correlation (xc) functional (especially in the case
(BSE),32 have yielded electronic properties in very good of solids)38,39 is its principal limitation.
agreement with experiments for a wide range of materials.33–39 The xc kernel (fxc ) of TDDFT, which is the variation of the
In the MBPT scheme, the QP wave functions ϕi and energies xc potential with respect to the electronic density, is most often
εi satisfy the Schrödinger-like equation approximated by (i) fxc = 0, leading to the random-phase
approximation (RPA), or (ii) fxc = δvxc LDA
/δρ, which gives the
45,46
[−∇ 2 /2 + Vext (r) + VH (r)]ϕi (r) adiabatic LDA (ALDA). Note that the RPA differs from
 an independent-particle calculation, even if the xc kernel is
+ (r,r ; εi )ϕi (r )d 3 r  = εi ϕi (r) , (1) 0, because it accounts for variations of the Hartree potential
upon excitation. This is a classical contribution, which, in the
where Vext is the external (pseudo)potential due to the ions, optical spectra, corresponds to local field or depolarization
VH is the classical Hartree potential, and the self-energy  effects, particularly important when charge inhomogeneities
contains all the effects of exchange and correlation among are pronounced.
electrons. Concerning the calculation of absorption spectra of semi-
In the GW approximation31 the self-energy is written as conductors, neither the RPA nor the ALDA is sufficient
the product to yield results in quantitative (and sometimes not even
in qualitative) agreement with experiments. Nevertheless, it
GW (1,2) = iG(1,2)W (1+ ,2), (2) has been shown39,47 that in some cases an agreement with
experiments can be recovered within TDDFT by retaining only
where G(1,2) is the one-particle Green’s function of the system the long-range contribution (LRC) that the exact fxc should
and W (1+ ,2) the dynamically screened Coulomb interaction. have in the asymptotic limit. This model kernel accounts
The argument 1 stands for (r1 ,σ1 ,t1 ) and 1+ means that the limit for continuum excitonic effects, giving very good optical
t1 + η → t1 should be taken for η positive. Usually, Eq. (1) is properties for small- to medium-gap semiconductors, with a
solved in the perturbative G0 W0 scheme, where G and W are computational burden comparable to a DFT approach.
calculated directly from LDA eigenvalues and eigenfunctions, Two forms for this LRC kernel are used in this work:
while the QP corrections to LDA eigenvalues are evaluated a static47,48 form given by −α static /q 2 and its dynamical
as a first-order perturbation. Applying the G0 W0 approach extension49 −(α dyn + βω2 )/q 2 , where α static , α dyn , and β are
gives very good results for many materials.34 However, it has parameters that can be calculated from the knowledge of the
been shown that it fails for compounds with d electrons,35,36 static dielectric constant and the plasma frequency of the
especially when d states hybridize with p states close to the material.48,49 One should remember, however, that CuGaS2
Fermi energy.6,40–42 When G0 W0 fails, a self-consistent (sc) possesses bound excitons10,21,26,27 and that the TDDFT-LRC
scheme or better starting points are the possible ways to obtain approach has proved to work better for systems with continuum
accurate electronic bands. excitons. That is why in this case it is worth resorting

085145-2
FIRST-PRINCIPLES STUDY OF THE BAND STRUCTURE . . . PHYSICAL REVIEW B 84, 085145 (2011)

to solution of the computationally demanding BSE, which to the core. For this reason, GW corrections can contain a
accurately describes the electron-hole pair dynamics. Indeed, substantial error when the spatial overlap between the valence
applications of BSE to several materials have been reported in and the semicore wave functions is sizable. This fact led
the literature,38,39 with excellent agreement with experiments. in the past to some claims64 concerning the failure of the
In the following we present results for the electronic pseudopotential approach for GW calculations. However, it
band structure (Sec. III) and optical absorption spectra has been proved65–67 that when semicore states are included in
(Sec. IV) of CuGaS2 that we obtained by applying the the valence GW results are reliable.
above-described hierarchy of ab initio schemes. We carried In our calculations this problem appears, as d electrons have
out similar calculations for the other members of the CIGS to be taken into account explicitly in the pseudopotential, as
family and the results are strictly analogous, as a consequence in the case of Cu or Ga. In the case of Cu, it has been shown
of the fact that in all CIGS compounds the physics of the band that it is crucial to take the whole shell in the valence68,69 in
gap is controlled by the p-d hybridization of the top valence order not to obtain unphysical GW corrections. Marini et al.68
states. proved that this behavior is due to the fact that Cu 3s and 3p
On one hand, our results allow us to assess the validity states, despite being well separated in energy from the 3d ones,
of different approximations and calculation schemes for the have a large spatial overlap with the latter. As a consequence,
determination of electronic and optical properties of Cu-based non-negligible contributions to the self-energy come from the
chalcopyrite materials. Our aim is to determine how to exchange contributions between 3d and 3s-3p states. Similar
obtain sufficiently accurate results using the most efficient conclusions were reached by Bruneval et al.,69 who showed
approaches, in order to lay the basis for tackling, in the that if Cu semicore states are included in the core, the GW
near-future, the more complex problem of studying the optical approximation predicts wrongly Cu2 O to be metallic.
absorption of CuGaS2 doped with transition metals. One As a consequence of the selected core-valence partition,
should remember that scGW calculations and the solution the kinetic energy cutoff used to describe the electronic
of the BSE are too computationally demanding at present in density and the wave functions becomes rather high (between
order to be applied systematically for large supercells made of 80 and 130 Ha, depending on the choice made for the Ga
atoms with many valence electrons. pseudopotential). While the valence-band maximum (VBM)
Moreover, comparison of the most accurate calculations in CuGaS2 4 has a considerable Cu 3d character, the necessity
with available experimental data for CuGaS2 is used to for the inclusion of Ga 3d states in the valence is less
analyze the origin of spectral features and clarify the existing obvious due to their high binding energies (19.85 eV with
discrepancies among experimental results. Our conclusions respect to the VBM).70 Note, moreover, that the generation of
are summarized in Sec. V. reliable Ga pseudopotentials has been a challenge for many
years.71 To shed light on this issue, we tested two different
Ga pseudopotentials: one, which we refer to as the Ga pseu-
II. COMPUTATIONAL DETAILS
dopotential without a semicore, has three valence electrons
All electronic-structure calculations were performed using (4s 2 4p1 ), while the other also includes 3d 10 , 3s 2 , and 3p6
the first-principles code ABINIT.35,50–52 We first generated and electrons in the valence (Ga pseudopotential with semicore).
tested ab initio norm-conserving pseudopotentials,53 obtained To evaluate the effect of the inclusion of semicore states in
within the Troullier-Martins54 and the Hamann55 schemes the Ga pseudopotential, we calculated the band structure of
using the code FHI98PP.56 The electronic ground state was CuGaS2 and, in particular, the value of the band gap using the
calculated using DFT within the LDA approximation of two Ga pseudopotentials with a different number of valence
Perdew and Wang57 for the xc functional. LDA eigenstates electrons. We observed that the band dispersions were very
were used as a starting point for either G0 W0 or scCOHSEX similar, independently of which pseudopotential we used, and
calculations. In the case of G0 W0 , the convolution in the there were only discrepancies as small as 0.1 eV in the value of
Fourier transform of Eq. (2) was performed using a plasmon- the QP band gap yielded by G0 W0 , and scCOHSEX + G0 W0
pole model for the frequency dependence of the dielectric calculations. Only the strongly underestimated KS LDA band
function.58 gap showed a larger discrepancy (0.25 eV). We concluded
In this work we used the fully relaxed primitive cell and therefore that using the Ga pseudopotential without a semicore
the internal coordinates obtained within the DFT-LDA. The would not significantly affect the accuracy of our GW results.
relaxed structural parameters (a = 5.3819 Å, c = 10.66 Å) More precisely, we estimate that our energy levels have an
are in agreement with experimental data within less than 2%. error bar of 0.1–0.2 eV, which is comparable to the difference
The relaxed anion displacement u is equal to 0.2581, compared between experimental results. In view of this, we decided to use
to an experimental range of [0.25–0.275].59–62 Note that in the in the following the Ga pseudopotential with only three valence
case of CuIn(S,Se)2 , the LDA and GGA gave a value of u electrons. This choice is particularly attractive for limiting the
outside the experimental window.6 We verified that this does computational effort as much as possible, especially in view
not happen when dealing with CuGaS2 and CuGaSe2 . of the fact that the application of CuGaS2 as an intermediate-
It is known63 that for GW calculations special care must band host requires the construction of supercells including
be taken in the choice of the pseudopotential core-valence doping.
partitions. When the core and valence electrons are separated The number of bands used in the sums over states for
in the procedure to build a pseudopotential, the self-energy the calculations of W and G was 250 and 200, respectively.
has a contribution from the core-valence exchange interaction, Furthermore, the following energy cutoffs of the plane-
which is treated at the LDA level if the semicore states belong wave basis were necessary to converge eigenvalues with a

085145-3
AGUILERA, VIDAL, WAHNÓN, REINING, AND BOTTI PHYSICAL REVIEW B 84, 085145 (2011)

7 full Brillouin zone. Moreover, 20 empty bands were needed

2744 k-points
to reach convergence in the range of photon energies from
6 1000 k-points 0 to 6 eV.
512 k-points The solution of the BSE is computationally heavier than
64 k-points
5 an RPA or ALDA calculation using TDDFT; for this reason,
we could not use a 14 × 14 × 14 shifted grid of k points.
4 A 10 × 10 × 10 shifted grid (which gives 1000 k points)
ε2(ω)

yields a spectrum where the relevant peaks are already well

3 defined, despite the presence of spurious wiggles due to the
finite k-point sampling. Since a similar convergence issue is
2 present in the RPA calculations, we show in Fig. 1 the slow
convergence with respect to the number of k points of the
1 RPA absorption spectrum. An analogous convergence issue
was already detected in previous works.73 In this case, it is
0 determined by the strong dispersion of the first conduction
1 2 3 4 5 6
Energy (eV) band close to . Nevertheless, one can see that a mesh
of 1000 shifted k points is clearly sufficient to determine
FIG. 1. (Color online) Convergence of the RPA absorption
the peak position with a precision of 0.1 eV. In fact, many
spectrum with respect to the density of the k-point grid.
approximations are involved in the various steps of our
calculations and we can expect an accuracy of the order of
tolerance of less than 20 meV for calculations with the Ga 0.1 eV in the position of our peaks. Once the peak position
pseudopotential without a semicore: 80 Ha for the KS wave and relative intensity of the peaks do not evolve any more (as
functions, 50 Ha for the wave functions used to calculate W , proven in Fig. 1) by increasing the number of k points, the
15 Ha for the size of the W matrix in reciprocal space, and 70 presence of some residual wiggles is acceptable, as long as
Ha for the wave functions used to calculate  and for the size one keeps their origin in mind.
of the x matrix in reciprocal space. For calculations with the
Ga pseudopotential with a semicore, note that the inclusion
III. QUASIPARTICLE BAND STRUCTURES
of the shell n = 3 in the valence of the Ga pseudopotential
leads to a significant increase in the kinetic energy cutoffs. The band gaps of CuGaS2 obtained using the different the-
Approximately a double number of plane waves was thus oretical approaches are summarized in Table I. As mentioned
necessary for convergence in the case of this pseudopotential. above, the standard G0 W0 approach fails for chalcopyrites
In the sc cycles, the COHSEX wave functions were because the DFT-LDA states are not a good starting point for a
represented on a restricted LDA basis set, with 59 and 42 LDA simple perturbative correction of the eigenenergies. As a result,
bands for calculations with the Ga pseudopotential with and the G0 W0 band gap (1.34 eV) is still far from the experimental
without a semicore, respectively. A Monkhorst-Pack k-point one (2.4–2.53 eV). When a perturbative G0 W0 fails, accurate
mesh of 3 × 3 × 3 was used to sample the Brillouin zone band gaps can be obtained moving away from the DFT starting
for GW calculations. This corresponds to 6 k points in the point, using, e.g., a restricted sc scheme. Self-consistency
irreducible wedge of the Brillouin zone and 27 k points in the using the COHSEX approximation for the self-energy leads, as
full Brillouin zone. usual, to a band gap that is too large (3.29 eV). The inclusion
Calculations of the optical spectra were performed using of dynamical effects through a perturbative G0 W0 step on
the code YAMBO.72 To get a well-converged RPA dielectric top of scCOHSEX corrects the result: the difference between
function, a 14 × 14 × 14 shifted grid was needed to sample experiment and scCOHSEX + G0 W0 is about 0.1 eV. In fact,
the Brillouin zone. This corresponds to 2744 k points in the we remind the reader that the experimental measurements

TABLE I. Theoretical and experimental band gaps Eg and binding energies of the valence states compared to XPS results. All energies are
given in electron volts with respect to the valence band maximum. Results for the Ga 3d binding energy are taken from the calculation done
using the Ga pseudopotential with a semicore.

LDA G0 W0 scCOHSEX scCOHSEX + G0 W0 Expt.

Eg 0.70 1.34 3.29 2.65 2.4−2.53a

Cu 3d + S 3p
First peak 1.5 1.4 1.8 1.7 1.6,2.1b
Second peak 3.2 3.3 3.6 3.6 3.4b
Ga-S bond 6.9 7.5 7.6 7.8 7.1b
S 3s 12.9 13.2 14.4 14.0 13.0,b 13.4c
Ga 3d 15.3−15.7 17.1−17.2 19.2−19.5 19.4−19.6 18.8 −19.3,b 19.85c
a
From Refs. 8, 9, and 12.
b
From Ref. 74.
c
From Ref. 70

085145-4
FIRST-PRINCIPLES STUDY OF THE BAND STRUCTURE . . . PHYSICAL REVIEW B 84, 085145 (2011)

yield optical gaps, which means that the difference with (a) (b)
respect to our calculated QP gap must be decreased by the 6 6 Conduction
excitonic binding energy of around 0.05 eV.11 We also remind 4 4 band
that the error due to the neglect of the semicore in the Ga 2 2
pseudopotential is of the same order. 0 0
In Table I we have also gathered the positions of the

E-E F (eV)
-2 -2 Cu−3d states

E-EF (eV)
main peaks of the density of states (DOS) calculated at the +
-4 -4
S−3p states
different theoretical levels, together with x-ray photoelectron -6 -6
spectroscopy (XPS) results from Refs. 70 and 74. The upper -8 -8 Ga−S bond
valence band is responsible for two peaks, corresponding to -10 -10
bonding and antibonding of hybridized Cu 3d and S 3p states. -12 -12
The following group of bands is made of Ga-S bonding states. -14 -14
S−3s states
Then we find S s states and, finally, the d states of Ga. -16 -16
T Γ N T Γ N
The width of both theoretical and experimental structures
has to be taken into account for the comparison. However, it FIG. 2. Band structure of CuGaS2 using (a) the LDA and (b)
can be clearly seen that a GW approach improves LDA results, scCOHSEX + G0 W0 .
especially when d states are involved. Indeed, for states with
d character (Cu 3d–S 3p hybridization and Ga 3d states),
we expect that LDA wave functions are too delocalized. In window close to the gap. In particular, we observed that the
particular, for the 3d states of Ga (which are obtained from the valence band width close to the gap changes by less than
calculations with the Ga pseudopotential with semicore states 0.1 eV, and analogously, the distance between the conduction
in the valence), the scGW approach shifts their position by valleys at , N , and T changes by less than 0.1 eV. We
4 eV, placing them very close to experimental data. In the case also evaluated the effective masses at  and compared them
of the Ga-S bonding and S 3s states, GW tends to increase the with previous calculations76 and experimental data.11 Our
binding energies in the right direction, but leading to a slight results, summarized in Table II, are in excellent agreement with
overcorrection. It should be noted that for the highest binding experiments. Note that we do not include spin-orbit coupling
energies, the predicted values cannot be considered as accurate in our calculations. However, it is known that its inclusion is
as for low binding energies, due to the use of the plasmon-pole essential to describe the dispersion of the top valence of Se
model.75 This issue does not affect the calculations of the compounds, while it has minor effects on S compounds.
optical spectra, as the energy region of interest is close to Note that, as in the case of CuInS2 ,6 the gap is strongly
the Fermi energy. dependent on the value of the anion displacement u: different
In Fig. 2 we show the dispersion of LDA and scCOHSEX + values of u within the experimental range [0.25–0.275] can
G0 W0 electronic energy bands of CuGaS2 . The correction to lead to variations of the band gap by 0.4 eV in LDA and
the band gap within scCOHSEX + G0 W0 is very large: the 0.8 eV in scCOHSEX + G0 W0 .
direct band gap at  goes from 0.7 eV in LDA to 2.65 eV, In order to verify that the QP wave functions and the LDA
as a result of a downshift of the valence band edge by 0.67 ones differ significantly, we analyzed the overlap between
eV and an upshift of the conduction band edge by 1.28 eV. LDA and COHSEX wave functions. This overlap is an
However, GW corrections are not very sensitive to the k point, indication of the variation of COHSEX wave functions with
and therefore the overall dispersion of the bands in LDA and respect to LDA ones. We found that there is a number of
scCOHSEX + G0 W0 is similar, which justifies the use of a states for which the overlap between scCOHSEX and LDA
scissor operator to simulate GW corrections in the energy wave functions is significantly smaller than 1 (down to 0.8),

TABLE II. Electron effective masses mc1 for the first conduction band at  and hole effective masses mv1,v2,v3 for the three highest valence
states at . All effective masses are given in units of electron mass.

Effective mass LDA scGW GGA + U from Ref. 76 Expt. from Ref. 11

m⊥c1 0.13 0.13 0.24

mc1 0.12 0.12 0.24
Average mc1 0.13 0.13 0.24 0.12
m⊥v1 1.22 0.90 0.91
mv1 0.16 0.15 0.30
Average mv1 0.87 0.65 0.70 0.68
m⊥v2 5.37 1.60 0.42
mv2 0.94 0.78 0.82
Average mv2 3.90 1.33 0.55 1.44
m⊥v3 0.16 0.16 0.41
mv3 0.94 0.78 0.81
Average mv3 0.42 0.36 0.54 0.39

085145-5
AGUILERA, VIDAL, WAHNÓN, REINING, AND BOTTI PHYSICAL REVIEW B 84, 085145 (2011)

CuGaSe2 CuInSe2 10
8 8
9 ALDA
6 6 LDA GW-RPA
scGW 8 exp. Levcenko
4 4 exp. Alonso
2 7
2
0
0 6

ε2(ω)
-2
E-EF (eV)

-2 5
-4
-4 -6 4

-6 -8 3
-8 -10 2
-10 -12
1
-12 -14
0
-16 1 2 3 4 5 6
-14 Energy (eV)
-18
-16
T Γ N T Γ N
FIG. 4. (Color online) Imaginary part of the dielectric function
for light polarized along the c axis obtained with TDDFT using the
FIG. 3. (Color online) Band structures of CuGaSe2 and CuInSe2
ALDA kernel [dashed (red) line], the RPA using GW eigenvalues
obtained using the LDA [dashed (red) lines] and scCOHSEX + G0 W0
[dot-dashed (black) line] compared to experimental data of Alonso
[solid (black) lines].
et al.7 (filled circles) and Levcenko et al.21 [solid (black) line; in
arbitrary units; see text].
signaling a pronounced change in the wave functions due to the
sc iterations. The projections are smaller for the more localized
states of the valence, which correspond to flatter bands. For the reflecting surface of the samples. In view of the above, in
these states, the performance of the LDA is worse and so all graphs we rescaled the intensity of the absorption spectrum
the QP wave functions differ more from the LDA ones. As a of Levcenko et al.21
matter of fact, the states at the top valence, formed by a p-d Another discrepancy among the available experimental data
hybridization, showed the smallest overlaps. concerns the behavior of the spectra at around 5 eV. While the
The results we have discussed for the band structure of measurements in Ref. 7 found a plateau in absorption after
CuGaS2 can be generalized to all members of the CIGS the peak at 4 eV, in Refs. 21 and 74 a deep valley appears
family, as the physics that determine the band gap is, in all after 4 eV. All calculations we performed found a behavior
these compounds, the repulsion of antibonding and bonding qualitatively similar to the data reported in Ref. 21.
p-d hybridized states. As we can see in Fig. 3, the scGW In Fig. 4 we compare results obtained within the RPA using
corrections to the LDA band structures of CuGaSe2 and scCOHSEX + G0 W0 eigenvalues77 and the ALDA using KS
CuInSe2 are, to a large extent, rigid shifts as in the case eigenvalues. The RPA curve obtained using KS eigenvalues
of CuGaS2 . The band gap correction is more spectacular in overlaps almost perfectly with the ALDA curve. This proves
CuInSe2 , as the LDA gap is negative for the In compounds of that the inclusion of the ALDA kernel does not have any
the CIGS family: in fact, the top valence band at  is located relevant effect on the absorption, as the important contributions
above the lowest conduction state. The correct band ordering due to electron-electron and electron-hole interactions are not
is re-established using a GW approach. accounted for by the ALDA kernel38,48,78 . Both the RPA and
the ALDA spectra using KS GGA eigenvalues are dramatically
red-shifted.
IV. OPTICAL PROPERTIES
We have also verified that an RPA spectrum that does
Starting from the KS and GW band structures presented not include local field effects is basically identical to an
in Sec. III, we calculated optical absorption spectra within RPA spectrum with local field effects. If we now use the
a linear response, applying TDDFT and solving the BSE. scCOHSEX + G0 W0 eigenvalues for an RPA calculation
Figures 4 and 5 show calculations of the imaginary part of [Fig. 4; dot-dashed (black) line], the resulting spectra are
the macroscopic dielectric function for light polarized along blue-shifted, while the line shape does not change. In fact, in
the c axis. The theoretical curves are also compared with the the energy range of interest for absorption the GW correction
experimental data published in Refs. 21 and 74. is basically equivalent to application of a scissor operator. If we
The experimental dielectric functions of Alonso et al.7 were do a comparison with the experimental curves, the absorption
obtained with ellipsometry at room temperature, whereas the peaks are now at energies that are too high. Additionally, the
spectra of Levcenko et al.21 were extracted from reflection intensity of the RPA spectrum just above the onset appears
spectra measured at 77 K, using the Kramers-Kronig relations. underestimated. This is due to the neglect of excitonic effects
Both results are for single-crystal samples. We can observe that within this approximation.
the two curves are rather different. In particular, the absorption For many intermediate-gap semiconductors, it was
spectrum of Levcenko et al.21 has a much lower intensity shown47–49 that the LRC to the xc kernel was enough to yield
(about half) than the experimental curve of Alonso et al. and a good agreement with experimental absorption spectra. The
previous results of Rife et al.74 , probably due to the quality of LRC kernel works particularly well in the case of a continuum

085145-6
FIRST-PRINCIPLES STUDY OF THE BAND STRUCTURE . . . PHYSICAL REVIEW B 84, 085145 (2011)

10 Of course, if one wants a reliable description of the bound

9 TDDFT static LRC
excitonic peak of CuGaS2 ,10,21,26,27 it is advisable to resort to
TDDFT dynamic LRC the solution of the BSE. Despite the fact that a shifted mesh
8 GW-RPA
BSE of 1000 k points is not sufficiently dense to smooth out the
exp. Levcenko
7
exp. Alonso
curve, due to the strong dispersion of the lowest conduction
6
state, the onset (at about 2.55 eV) and the peak (at about
3.8 eV) are already well converged (see Fig. 1) and they are in
ε2(ω)

5 excellent agreement with experiments. At the absorption edge,

4 we estimate an excitonic binding energy of about 0.1 eV, in
agreement with experimental data.11
3
The peak at 3.8 eV is a resonant exciton coming from
2 transitions between the two last valence bands and the first
1 conduction band at k points in the region around the symmetry
point N. The flatness of the bands in that region explains the
0
1 2 3 4 5 6 large effect that the BSE has on this peak.
Energy (eV)
By comparing, in Fig. 6, the absorption for light polarized
FIG. 5. (Color online) Imaginary part of the dielectric function perpendicularly to the c axis or along the c axis, we can observe
for light polarized along the c axis obtained with the TDDFT using the that the two experiments are once again in disagreement
static [dot-dashed (black) line] and the dynamic [dashed (red) line] concerning the intensity of the absorption in the energy region
LRC kernel, the GW-RPA [thin solid (gray) line], and the Bethe- of the peak at about 4 eV. All our calculations agree better in
Salpeter equation [dotted (blue) line], compared to experimental this case with the data of Alonso et al.7 and predict a higher
measurements of Alonso et al.7 (filled circles) and Levcenko et al.21 absorption for light polarized in the plane perpendicular to
[thick solid (black) line; in arbitrary units]. the c axis. The anisotropy of the optical response reflects
an anisotropy of the band structure that is already present
exciton, giving, in general, very good optical properties with in the RPA spectra and is not significantly modified in
the efficiency of a TDDFT calculation. In Fig. 5 we show the the BSE result. The disagreement with the intensity of the
spectra we obtained using the static47,48 and the dynamic49 polarization components of the spectra of Levcenko et al.21
versions of the LRC kernel. scCOHSEX + G0 W0 eigenvalues may, however, come from additional effects due to the way
were used in these calculations. their absorption spectra were obtained. First, the polarization
We extracted the parameter α static required to define the components when one measures reflectance may depend on the
static LRC kernel from theoretical quantities. More precisely, plane of growth of the samples [reflectances in Ref. 21 were
α static = 0.37 was obtained from the expression in Ref. 48 measured in the (110) plane]. Second, it is well known80,81
using the RPA value of the dielectric constant ε∞ (7.9243). that extrapolation of the reflectance to the high-energy region
For the dynamical LRC, the expression in Ref. 49 leads to (outside the range of the measurements) in order to apply
values of α dyn = 0.11 and β = 0.004. However, CuGaS2 has a the Kramers-Kronig relations affects the absolute value of the
bound exciton with strong effects on the spectrum, and in that optical spectra. Finally, both experiments and our calculations
case, the formula in Ref. 49 for the prediction of α dyn and β agree that when light is polarized along the c axis, the
may not be accurate. In fact, in order to obtain a reasonable absorption onset is at a slightly lower energy and the first
spectrum for CuGaS2 , we need to use values of α dyn = 0.11 excitonic peak is more pronounced.
and β = 0.02, resorting to an empirical approach. The fact that Note that two other effects can contribute to the differences
local field effects are small gives us a hint of why the LRC between experimental and theoretical onsets: the finite tem-
kernel cannot modify much the spectrum.79 Moreover, as the perature of the experimental setup and, in the case of Ref. 7, a
valence states involved are rather flat, one can expect that the non-negligible amount of In in the samples. Both effects tend
LRC approximation is not enough to describe excitonic effects. to close band gaps.
The TDDFT-LRC spectra shown in Fig. 5 provide a clear
improvement with respect to the curves displayed in Fig. 4.
V. SUMMARY AND CONCLUSIONS
The intensity of the peaks as well as the onsets of absorption
are now closer to experiment. However, the peak at about We used a state-of-the-art QPscGW approach to deter-
4 eV is still located at a higher energy than its experimental mine the electronic properties of CuGaS2 , which has been
position. Note that the excitonic peak at the fundamental edge recognized as an ideal host semiconductor for intermediate-
in CuGaS2 has the nature of a bound exciton, even if the band thin-film solar cells. Perturbative GW is known to fail
binding energy is rather low. This explains48 why the LRC for Cu-based chalcopyrites, due to the p-d hybridization of the
models do not reproduce the Bethe-Salpeter result perfectly. states forming the top of the valence band.6 Indeed, this is also
However, one should not forget that a TDDFT calculation true for CuGaS2 : we found a G0 W0 band gap almost 50% too
using a model xc kernel is much less involved than a calculation small. An accurate band gap, in agreement with experimental
which requires the solution of the BSE. In many cases, for data within 0.1 eV, was obtained using scCOHSEX + G0 W0 .
example, when dealing with supercells with dopants, the best In the energy range of interest for optical absorption the GW
possible compromise, in terms of accuracy and computational corrections shift the KS LDA bands almost rigidly, as we
costs, is clearly to perform calculations using model kernels proved through the analysis of the effective masses, band
derived from Bethe-Salpeter. widths, and relative positions of the conduction energy valleys.

085145-7
AGUILERA, VIDAL, WAHNÓN, REINING, AND BOTTI PHYSICAL REVIEW B 84, 085145 (2011)

30 KS eigenvalues, yield a dramatic red-shift of the absorption

⊥
spectra with respect to experiments, which reflects the strong
|| underestimation of the gap in KS. The use of GW eigenvalues
25 Exp. Alonso et al.
for a subsequent RPA calculation leads to spectra that are
blue-shifted compared to experiments. This is due to the
20
complete neglect of excitonic effects.
The inclusion of excitonic effects via the static and dy-
Absorption

Exp. Levcenko et al.

15 namical LRC kernel of the TDDFT improves the results: both
the intensities of the peaks and the position of the absorption
10 onset are significantly closer to experiments. However, the
peak at about 4 eV still appears at higher energies than its
5 BSE calculation experimental position and the bound excitonic peak at the
RPA calculation fundamental edge is not reproduced. Only the solution of the
BSE gives an accurate quantitative description of absorption.
0
1 2 3 4 5 6 Nevertheless, we can conclude that the TDDFT can be an
Energy (eV)
efficient and reliable alternative to the BSE, especially in view
FIG. 6. (Color online) Comparison of BSE and GW-RPA absorp- of calculations for large supercells, when the model LRC
tion spectra with experimental absorption spectra from Refs. 7 and 21. xc kernels derived from the BSEs are employed. Also, this
Dashed (black) and solid (red) curves are results for light polarization conclusion is extremely relevant for further ab initio studies
perpendicular and parallel to the c axis, respectively. of doped systems, which require the use of large supercells,
in order to describe the effect of the intermediate band on the
This finding justifies the use of a scissor operator to correct the optical absorption.
band gap. However, one should remember that the QP and LDA
wave functions differ significantly for valence states close to
ACKNOWLEDGMENTS
the band gap, which are composed of hybridized s-p states.
We tested the importance of the inclusion of d states in the This work was supported by the Consolider Ingenio
pseudopotential of Ga. We found that the differences in the 2010 Program, through project GENESIS-FV (reference no.
GW gap are as small as 0.1 eV when including or excluding CSD2006-04) of The Spanish Ministry of Education and
the 3 shell of Ga in the valence, which is the precision we Science, the NUMANCIA 2 project (S-2009ENE-1477) of
can expect from a pseudopotential approach. The accuracy of the Comunidad de Madrid, and the EU Seventh Framework
GW results using the Ga pseudopotential without a semicore Programme under Grant Agreement No. 211956. The authors
is then of the order of the difference between experimental thankfully acknowledge the computer resources, technical
results. This is important for reducing the computational costs expertise, and assistance provided by the Centro de Super-
in view of future calculations for CuGaS2 doped with transition computación y Visualización de Madrid (CeSViMa) and the
metals. Spanish Supercomputing Network. Some of the calculations
Regarding optical properties, the accuracy of RPA or ALDA were performed at IDRIS (Project Nos. x2010096017 and
ab initio calculations was not sufficient to explain the large 09 544). I.A. would also like to thank the FOTOMAT
discrepancies among experimental optical measurements. In project (MAT2009-14625-C03-01) for a research contract.
view of this, we compared optical properties obtained with S.B. acknowledges financial support from Programme PIR
several theoretical levels of approximations. The simplest Matériaux–MaProSu No. 55744 of the CNRS. J.V. acknowl-
approaches, RPA or TDDFT with an ALDA kernel, using edges support from the EDF/ANR CIFRE.

* 8
silvana.botti@polytechnique.edu J. L. Shay, B. Tell, H. M. Kasper, and L. M. Schiavone, Phys. Rev.
1
I. Repins, M. A. Contreras, B. Egaas, C. DeHart, J. Scharf, B 5, 5003 (1972).
9
C. L. Perkins, B. To, and R. Noufi, Prog. Photovolt. 16, 235 B. Tell, J. L. Shay, and H. M. Kasper, Phys. Rev. B 4, 2463 (1971).
10
(2008). C. Bellabarba, J. González, and C. Rincón, Phys. Rev. B 53, 7792
2
P. Jackson, D. Hariskos, E. Lotter, S. Paetel, R. Würz, R. Menner, (1996).
11
W. Wischmann, and M. Powalla, Prog. Photovolt. Res. Appl., in N. N. Syrbu, I. M. Tiginyanu, L. L. Nemerenco, V. V. Ursaki,
press (2011), doi:10.1002/pip.1078. V. E. Tezlevan, and V. V. Zalamai, J. Phys. Chem. Solids 66, 1974
3
J. E. Jaffe and A. Zunger, Phys. Rev. B 29, 1882 (1984). (2005).
4 12
J. E. Jaffe and A. Zunger, Phys. Rev. B 27, 5176 (1983). J. R. Botha, M. S. Branch, P. R. Berndt, A. W. R. Leitch, and
5
F. D. Jiang and J. Y. Feng, Semicond. Sci. Technol. 23, 025001 J. Weber, Thin Solid Films 515, 6246 (2007).
13
(2008). W. Shockley and H. J. Queisser, J. Appl. Phys. 32, 510 (1961).
6 14
J. Vidal, S. Botti, P. Olsson, J.-F. Guillemoles, and L. Reining, Phys. D. F. Marrón, A. Martı́, and A. Luque, Thin Solid Films 517, 2452
Rev. Lett. 104, 056401 (2010). (2009).
7 15
M. I. Alonso, K. Wakita, J. Pascual, M. Garriga, and N. Yamamoto, I. Aguilera, P. Palacios, and P. Wahnón, Thin Solid Films 516, 7055
Phys. Rev. B 63, 075203 (2001). (2008).

085145-8
FIRST-PRINCIPLES STUDY OF THE BAND STRUCTURE . . . PHYSICAL REVIEW B 84, 085145 (2011)

16 50
P. Palacios, I. Aguilera, and P. Wahnón, Thin Solid Films 516, 7070 ABINITcode; [www.abinit.org].
51
(2008). X. Gonze, J.-M. Beuken, R. Caracas, F. Detraux, M. Fuchs, G.-M.
17
P. Palacios, I. Aguilera, P. Wahnón, and J. C. Conesa, J. Phys. Chem. Rignanese, L. Sindic, M. Verstraete, G. Zerah, F. Jollet, M. Torrent,
C 112, 9525 (2008). A. Roy, M. Mikami, P. Ghosez, J.-Y. Raty, and D. Allan, Comput.
18
A. Luque and A. Martı́, Phys. Rev. Lett. 78, 5014 (1997). Mater. Sci. 25, 478 (2002).
19 52
A. Antolı́n, A. Martı́, J. Olea, D. Pastor, G. González-Dı́az, I. Mártil, X. Gonze, G.-M. Rignanese, M. Verstraete, J.-M. Beuken,
and A. Luque, Appl. Phys. Lett. 94, 042115 (2009). Y. Pouillon, R. Caracas, F. Jollet, M. Torrent, G. Zerah, M. Mikami,
20
A. Martı́, D. F. Marrón, and A. Luque, J. Appl. Phys. 103, 073706 P. Ghosez, M. Veithen, J.-Y. Raty, V. Olevano, F. Bruneval,
(2008). L. Reining, R. Godby, G. Onida, D. Hamann, and D. Allan,
21
S. Levcenko, N. N. Syrbu, V. E. Tezlevan, E. Arushanov, Z. Kristallogr. 220, 558 (2005).
53
S. Doka-Yamigno, T. Schedel-Niedrig, and M. C. Lux-Steiner, G. B. Bachelet, D. R. Hamann, and M. Schlüter, Phys. Rev. B 26,
J. Phys. Condens. Matter 19, 456222 (2007). 4199 (1982).
22 54
C. Guillén and J. Herrero, Thin Solid Films 515, 5917 (2007). N. Troullier and J. L. Martins, Phys. Rev. B 43, 1993 (1991).
23 55
S. Laksari, A. Chahed, N. Abbouni, O. Benhelal, and B. Abbar, D. R. Hamann, Phys. Rev. B 40, 2980 (1989).
56
Comput. Mater. Sci. 38, 223 (2006). M. Fuchs and M. Scheffler, Comput. Phys. Commun. 119, 67
24
R. Ahuja, S. Auluck, O. Eriksson, J. M. Wills, and B. Johansson, (1999).
57
Solar Energy Mater. Solar Cells 53, 357 (1998). J. P. Perdew and Y. Wang, Phys. Rev. B 45, 13244 (1992).
25 58
A. Soni, V. Gupta, C. Arora, A. Dashora, and B. Ahuja, Solar Energy R. W. Godby and R. J. Needs, Phys. Rev. Lett. 62, 1169 (1989).
59
84, 1481 (2010). J. Schneider, A. Rauber, and G. Brandt, J. Phys. Chem. Solids 34,
26
B. Tell and H. M. Kasper, Phys. Rev. B 7, 740 (1973). 443 (1973).
27 60
J. Ringeissen, J. L. Regolini, and S. Lewonczuk, Surface Sci. 37, S. C. Abrahams and J. L. Bernstein, J. Chem. Phys. 59, 5415 (1973).
61
777 (1973). H. Hahn, G. Frank, W. Klingler, A. Meyer, and G. Storger, Z. Anorg.
28
S.-H. Wei and A. Zunger, Phys. Rev. B 37, 8958 (1988). Chem. 271, 153 (1953).
29 62
P. Hohenberg and W. Kohn, Phys. Rev. B 136, 864 (1964). H. W. Spiess, V. Haeberln, G. Brandt, A. Rauber, and J. Schneider,
30
W. Kohn and L. J. Sham, Phys. Rev. A 140, 1133 (1965). Phys. Status Solidi B 62, 183 (1974).
31 63
L. Hedin, Phys. Rev. A 139, 796 (1965). M. S. Hybertsen and S. G. Louie, Phys. Rev. B 34, 5390 (1986).
32 64
E. E. Salpeter and H. A. Bethe, Phys. Rev. 84, 1232 (1951). W. Ku and A. G. Eguiluz, Phys. Rev. Lett. 89, 126401 (2002).
33 65
F. Aryasetiawan and O. Gunnarsson, Rep. Prog. Phys. 61, 237 M. L. Tiago, S. Ismail-Beigi, and S. G. Louie, Phys. Rev. B 69,
(1998). 125212 (2004).
34 66
W. G. Aulbur, L. Jönsson, and J. W. Wilkins, Solid State Phys. 54, K. Delaney, P. Garcia-Gonzalez, A. Rubio, P. Rinke, and R. W.
1 (1999). Godby, Phys. Rev. Lett. 93, 249701 (2004).
35 67
F. Bruneval, N. Vast, and L. Reining, Phys. Rev. B 74, 045102 C. Friedrich, A. Schindlmayr, S. Blügel, and T. Kotani, Phys. Rev.
(2006). B 74, 045104 (2006).
36 68
M. van Schilfgaarde, T. Kotani, and S. V. Faleev, Phys. Rev. Lett. A. Marini, G. Onida, and R. Del Sole, Phys. Rev. B 64, 195125
96, 226402 (2006). (2001).
37 69
T. Kotani, M. van Schilfgaarde, and S. V. Faleev, Phys. Rev. B 76, F. Bruneval, N. Vast, L. Reining, M. Izquierdo, F. Sirotti, and
165106 (2007). N. Barrett, Phys. Rev. Lett. 97, 267601 (2006).
38 70
G. Onida, L. Reining, and A. Rubio, Rev. Mod. Phys. 74, 601 E. P. Domashevskaya, L. N. Marshakova, V. A. Terenkhov, A. N.
(2002), and refereces therein. Lukin, Y. A. Ugai, V. I. Nefedov, and Y. Y. Salyn, Phys. Status
39
S. Botti, A. Schindlmayr, R. Del Sole, and L. Reining, Rep. Prog. Solidi B 106, 429 (1981).
71
Phys. 70, 357 (2007). O. A. von Lilienfeld and P. A. Schultz, Phys. Rev. B 77, 115202
40
M. Gatti, F. Bruneval, V. Olevano, and L. Reining, Phys. Rev. Lett. (2008).
72
99, 266402 (2007). A. Marini, C. Hogan, M. Grüning, and D. Varsano, Comput. Phys.
41
J. Vidal, F. Trani, F. Bruneval, M. A. L. Marques, and S. Botti, Phys. Commun. 180, 1392 (2009).
73
Rev. Lett. 104, 136401 (2010). M. Rohlfing and S. G. Louie, Phys. Rev. B 62, 4927 (2000).
42 74
F. Trani, J. Vidal, S. Botti, and M. A. L. Marques, Phys. Rev. B 82, J. C. Rife, R. N. Dexter, P. M. Bridenbaugh, and B. W. Veal, Phys.
085115 (2010). Rev. B 16, 4491 (1977).
43 75
S. V. Faleev, M. van Schilfgaarde, and T. Kotani, Phys. Rev. Lett. A. Marini, G. Onida, and R. Del Sole, Phys. Rev. Lett. 88, 016403
93, 126406 (2004). (2001).
44 76
E. Runge and E. K. U. Gross, Phys. Rev. Lett. 52, 997 (1984). C. Persson, Appl. Phys. Lett. 93, 072106 (2008).
45 77
E. K. U. Gross, F. J. Dobson, and M. Petersilka, Density Functional In practice, we verified that we can correct KS-GGA eigenvalues
Theory (Springer, New York, 1996). with a scissor operator of 1.7 eV. This allows us to match the
46
A. Zangwill and P. Soven, Phys. Rev. A 21, 1561 (1980). scCOHSEX + G0 W0 band structure in the energy range of interest
47
L. Reining, V. Olevano, A. Rubio, and G. Onida, Phys. Rev. Lett. for optical absorption.
78
88, 066404 (2002). V. I. Gavrilenko and F. Bechstedt, Phys. Rev. B 55, 4343 (1997).
48 79
S. Botti, F. Sottile, N. Vast, V. Olevano, L. Reining, H. C. Weissker, F. Sottile, F. Bruneval, A. G. Marinopoulos, L. K. Dash, S. Botti,
A. Rubio, G. Onida, R. Del Sole, and R. W. Godby, Phys. Rev. B V. Olevano, N. Vast, A. Rubio, and L. Reining, Int. J. Quantum
69, 155112 (2004). Chem. 102, 684 (2005).
49 80
S. Botti, A. Fourreau, F. Nguyen, Y.-O. Renault, F. Sottile, and G. Leveque, J. Phys. C 10, 4877 (1977).
81
L. Reining, Phys. Rev. B 72, 125203 (2005). K. Jezierski, J. Phys. C 17, 475 (1984).

085145-9