Computational Materials Science 28 (2003) 259273 www.elsevier.

com/locate/commatsci

Solid state calculations using WIEN2k

Karlheinz Schwarz *, Peter Blaha
Institute for Materials Chemistry, Vienna University of Technology, Getreidemarkt 9/165-TC, A-1060 Vienna, Austria

Abstract To study solid materials on the atomic scale one often starts with an ideal crystal at zero temperature and calculates its electronic structure by means of density functional theory (DFT). This allows a quantum mechanical treatment of the physics that underlines properties such as relative stability, chemical bonding, relaxation of the atoms, phase transitions, electrical, mechanical, optical or magnetic behavior, etc. For the solution of the DFT equations several methods have been developed. The linearized-augmented-plane-wave method is one of the most accurate methods. It is embodied in the computer codeWIEN2kwhich is now used worldwide by more than 500 groups to solve crystal properties on the atomic scale (see www.wien2k.at). Nowadays calculations of this type can be doneon suciently powerful computersfor systems containing about 100 atoms per unit cell. Chromium dioxide CrO2 is selected as a representative example using both, bulk and surface structures. References to other applications are given. 2003 Elsevier B.V. All rights reserved.
PACS: 71.10; 71.15; 71.20; 71.22 Keywords: WIEN2k; Density functional theory (DFT); Electronic structure

1. Introduction In advanced solid materials, which are of great technological interest, several parameters such as temperature, pressure, structure, composition, and disorder determine their properties. They are governed by very dierent length and time scales which may vary by many orders of magnitude depending on their applications. Let us focus on the length scale, where from meters (m) down to micrometers (lm) classical mechanics and continuum models are the dominating concepts to in-

Corresponding author. E-mail address: kschwarz@theochem.tuwien.ac.at (K. Schwarz). URL: http://info.tuwien.ac.at/theochem/.

vestigate the properties of the corresponding materials. However, when one comes to the nanometer (nm) scale or atomic dimensions measured  in A, the properties are determined by the electronic structure of the solid. In the development of modern materials an understanding on the atomic scale is frequently essential in order to replace trial and error procedures by a systematic materials design. Modern devices in the electronic industry provide such an example, where the increased miniaturization is one of the key advances. Other applications are found in the area of magnetic recording or other storage media. In chemistry heterogeneous catalysis is a typical example, in which one likes to understand the details of catalytic processes between molecules interacting with a solid surface as for example in a zeolite.

0927-0256/$ - see front matter 2003 Elsevier B.V. All rights reserved. doi:10.1016/S0927-0256(03)00112-5

260

K. Schwarz, P. Blaha / Computational Materials Science 28 (2003) 259273

One possibility to study complex systems is to perform computer simulations. Calculation of solids in general (metals, insulators, semiconductors, minerals, etc.) can be performed with a variety of methods from classical to quantum mechanical approaches. The former are force eld schemes, in which the forces that determine the interactions between the atoms are parameterized in order to reproduce a series of experimental data such as equilibrium geometries, bulk moduli or special vibrational frequencies (phonons). These schemes have reached a high level of sophistication and are often used within a given class of materials provided good parameters are already known from closely related systems. If, however, such parameters are not available, or if a system shows unusual phenomena that are not yet understood, one often must rely on ab initio calculations. These are more demanding in terms of computer requirements and thus allow only the treatment of smaller unit cells than force-eld calculations. The advantage of rst-principle (ab initio) methods lies in the fact that they can be carried out without any experimental knowledge of the system. In the following we will restrict ourselves to ab initio methods and sketch their main characteristics. The fact that electrons are indistinguishable and are Fermions requires that their wave functions must be anti-symmetric when two electrons are interchanged. This situation leads to the phenomenon of exchange. There are two types of approaches for a full quantum mechanical treatment, HF and DFT. The traditional scheme is (or was) the Hartree Fock (HF) method which is based on a wave function description (with one Slater determinant). Exchange is treated exactly but correlation eects are ignored by denition. The latter can be included by more sophisticated approaches such as the conguration interaction scheme that progressively requires more computer time with a scaling as bad as N 7 when the system size (N ) grows. As a consequence it is only feasible to study small systems, which contain a few atoms. An alternative scheme is density functional theory (DFT) that is commonly used to calculate the electronic structure of complex systems containing many atoms such as large molecules or solids. It is based on the electron density rather

than on wave functions and treats exchange and correlation, but both approximately. It will be discussed in the next section. The ideal crystal is dened by the unit cell which may contain several (up to about 100) atoms. Periodic boundary conditions are used to describe the innite crystal by knowing the properties in one unit cell. Translational and point group symmetry operations (inversion, rotation, mirror planes, etc.) that leave the ideal crystal invariant allow to simplify the calculations, which all correspond to the absolute temperature zero (T 0). 2. Density functional theory 2.1. The KohnSham equation The well-established scheme to calculate electronic properties of solids is based on DFT, for which Walter Kohn has received the Nobel Prize in chemistry in 1998. DFT is a universal approach to the quantum mechanical many-body problem, where the system of interacting electrons is mapped in a unique manner onto an eective noninteracting system with the same total density. Hohenberg and Kohn [1] have shown that the electron density . uniquely denes the total energy Etot of a system and is a functional Etot q of the density: Etot q Ts q Eee q ENe q Exc q ENN : 1 The dierent electronic contributions are conventionally labeled as, respectively, the kinetic energy (of the non-interacting particles), the electron electron repulsion, nuclearelectron attraction, and exchange-correlation energies. The last term corresponds to the repulsive Coulomb energy of the xed nuclei ENN . The most ecient way to minimize Etot by means of the variational principle is to introduce orbitals vik constrained to construct the densities as a sum over occupied orbitals according to the aufbau principle (with increasing energy) X qr qik jvik rj2 : 2
i;k

K. Schwarz, P. Blaha / Computational Materials Science 28 (2003) 259273

261

Here, the qik are occupation numbers such that 0 6 qik 6 1=wk , where wk is the symmetry-required weight of point k. Then variation of Etot gives a set of eective one-particle Schrdinger equations o (written for an atom in Rydberg atomic units), the so-called KohnSham equations [2], r VNe Vee Vxc vik r ik vik r;
2

reversed. This perspective illustrates the importance of improving the functional in DFT calculations, since this denes the quality of the calculation. 2.2. Spin polarization and magnetism Magnetism can come from localized electrons (e.g. from the 4f electron in rare earth elements) or from delocalized electrons. In the latter case the itinerant electrons can be well described by band theory. Without spinorbit coupling the magnetism originates from the spin polarization, whereas the orbital moment is often quenched. For example, Fe, Co, Ni are ferromagnets (FM) which fall in this category. In this case DFT is generalized to allow for a spin polarization, in which the spin-up and spin-down electrons move in spin-dependent potentials, which can be deduced from the corresponding spin densities, for example using the local spin density approximation. Another case are antiferromagnets (AFM) (e.g. bcc Cr) in which neighboring atoms have the opposite spin alignment but the magnitude of the moment at each site is the same. The net bulk magnetization vanishes for AFM. When the magnitude diers we have ferrimagnetism and a net magnetization. All these cases belong to collinear magnets. The orientation of the magnetic moment with respect to the crystal axis is only dened when spinorbit coupling is included. There are more complicated magnetic structures such as canted magnetic moments or spin spirals. In such cases the direction of the magnetization varies with position and this is called non-collinear magnetism. 2.3. Choice of basis set, wave function, and potential Many methods are available to solve the DFT equations. Essentially all use a linear combination of basis functions. Some use a linear combination of atomic orbitals (LCAO) method with Gaussian or Slater type orbitals (GTOs or STOs) others use mun tin orbitals (MTOs) as in linear combination of MTOs or augmented spherical wave. In the former cases the basis functions are given in analytic form, but in the latter the radial wave functions are obtained by numerically integrating

which must be solved. The non-interacting particles of this auxiliary system move in an eective local one-particle potential, which consists of a classical mean-eld (Hartree) part and an exchange-correlation part Vxc (due to quantum mechanics) that, in principle, incorporates all correlation eects exactly. The KS equations must be solved in an iterative process till self-consistency is reached, since nding the KS orbitals requires the knowledge of the potentials which themselves depend on the density and thus on the orbitals again. The exact functional form of Exc or the potential Vxc is not known and thus one needs to make approximations. Early applications were done by using results from quantum Monte Carlo calculations for the homogeneous electron gas, for which the problem of exchange and correlation can be solved numerically with high accuracy, leading to the original local density approximation (LDA). LDA works reasonably well but has some shortcomings mostly due to the tendency of overbinding, which cause e.g. too small lattice constants. Modern versions of DFT, especially those using the generalized gradient approximation (GGA), improve the LDA by adding gradient terms of the electron density and reach (almost) chemical accuracy. At present the version by PerdewBurke Ernzerhof is the recommended option for solids [3]. In the study of large systems the strategy of schemes based on HF or DFT dier. In HF based methods the Hamiltonian is well dened but can be solved only approximately. In DFT, however, one must rst choose (approximate) the functional that is used to represent the exchange and correlation eects but then this eective Hamiltonian can be solved almost exactly, i.e. with very high numerical accuracy. Thus the approximation necessarily enters in both cases but the sequence is

262

K. Schwarz, P. Blaha / Computational Materials Science 28 (2003) 259273

the radial Schrdinger equation. Alternatively o plane waves (PWs) form in principle a complete basis set. However, convergence is slow and thus one must use either a pseudo-potential or augment them with some other functions. Usually this augmentation is done with some atomic-like basis set. In the mun tin or the atomic sphere approximation (MTA or ASA) an atomic sphere, in which the potential (and charge density) is assumed to be spherically symmetric, surrounds each atom in the crystal. While these approximations work reasonably well in highly coordinated, closely packed systems (as for example face centered cubic metals) they become very approximate in all non-isotropic cases (e.g. layered compounds, semiconductors, or other open structures). Schemes that make no shape approximation in the form of the potential are termed full-potential schemes (see Section 3.2). Closely related to the basis sets is the explicit form of the wave functions, which should be well representable. These can be node-less pseudo-wavefunctions or all-electron wave-functions including the complete radial nodal structure and a proper description close to the nucleus. With a proper choice of pseudo-potential one can focus on the valence electrons, which are relevant for chemical bonding, and replace the inner part of their wave functions by node-less pseudofunctions that can be expanded in PWs with good convergence. 2.4. Choice of method As a consequence of the aspects described above dierent methods have their advantages or disadvantages when it comes to compute various quantities. For example, properties that rely on the knowledge of the density close to the nucleus (hyperne elds, electric eld gradients, etc.), require an all-electron description rather than a pseudo-potential approach with un-physical wave functions near the nucleus. On the other hand for studies, in which the shape (and symmetry) of the unit cell changes, the knowledge of the corresponding stress tensor is needed for an ecient structural optimization. These tensors are much easier to obtain in pseudo-potential schemes and

thus are available there. In augmentation schemes such algorithms become more tedious and consequently are often not implemented. On the other hand all-electron methods do not depend on choices of pseudo-potentials and contain the full wave function information. Thus the choice of method for a particular application depends on the property of interest and may aect the accuracy, ease or diculty to calculate a given property.

3. The linearized-augmented-plane-wave method One among the most accurate schemes for solving the KohnSham equations (see Section 2) is the full-potential linearized-augmented-planewave (FP-LAPW) method suggested by Andersen [4] on which our WIEN code is based. For further details see e.g. the book by Singh [5] with many references and details. During the last two decades this FP-LAPW code has been developed in our group and is used worldwide by more than 500 groups at universities and industrial laboratories. The original version was the rst LAPW code that was published and thus was made available for other users (WIEN) [6]. In the LAPW method [4] a basis set is introduced that is especially adapted to the problem by dividing the unit cell into (I) non-overlapping atomic spheres (centered at the atomic sites) and (II) an interstitial region (Fig. 1). For the construction of basis functionsbut only for that purposethe MTA is used according to which the potential is assumed to be spherically symmetric within the atomic spheres but constant outside. In

I I II
Fig. 1. Partitioning of the unit cell into atomic spheres (I) and an interstitial region (II).

K. Schwarz, P. Blaha / Computational Materials Science 28 (2003) 259273

263

the two types of regions dierent basis sets are used. (I) inside atomic sphere t, of radius Rt , a linear combination of radial functions times spherical harmonics Ylm r is used (we omit the index t when it is clear from the context) X _ /kn Alm;kn ul r; El Blm;kn ul r; El Ylm ^; r
lm

4 where ul r; El is the (at the origin) regular solution of the radial Schrdinger equation for energy El o (usually chosen at the center of bands with the _ corresponding -like character) and ul r; El is the energy derivative of ul evaluated at the same energy El . A linear combination of these two functions linearize the energy dependence of the radial function; the coecients Alm and Blm are functions of kn (see below) and are determined by requiring that this basis function matches (in value and slope) the PW labeled with kn , the corresponding basis function of the interstitial region. The func_ tions ul and ul are obtained by numerical integration of the radial Schrdinger equation on a o radial mesh inside sphere t within the spherical part of the potential. (II) in the interstitial region a PW expansion is used 1 /kn p eikn r ; x 5

In order to improve upon the linearization (i.e. to increase the exibility of the basis) and to make possible a consistent treatment of semicore and valence states in one energy window (to ensure orthogonality) additional (kn independent) basis functions can be added. They are called local orbitals (LO) according to Singh [7] and consist of a linear combination of two radial functions at two dierent energies (e.g. at the 3s and 4s energies) and one energy derivative (at one of these energies): _ /LO Alm ul r; E1;l Blm ul r; E1;l lm Clm ul r; E2;l Ylm ^: r 7

The coecients Alm , Blm and Clm are determined by the three requirements that /LO should be normalized and has zero value and slope at the sphere boundary. 3.1. Dierent implementations of APW The energy dependence of the atomic radial functions described above can be treated in different ways. In APW this is done by nding the energy that corresponds to the eigenenergy of each state. This leads to a non-linear eigenvalue problem, since the basis functions become energy dependent. In LAPW a linearization of the energy dependence is introduced by solving the radial Schrdinger equation for a xed linearization eno ergy but adding the energy derivative of this function [4,5]. The corresponding two coecients can be chosen such as to match (at the atomic sphere boundary) the atomic solution to each PW in value and slope, which determine the two coecients of the function and its derivative (see Eq. (4)). LAPW leads to a standard general eigenvalue problem but the PW basis is less ecient than in APW. A new scheme, APW plus local orbitals, [8] combines the advantages of both methods. As in LAPW a xed linearization energy is used but the energy dependence is covered by using a kn -independent local orbital. The matching is only done in value, but this new scheme leads to a signicant speed-up of the method (up to an order of magnitude) while keeping the high accuracy of LAPW [9]. A description of these three types of schemes

where kn k Kn ; Kn are the reciprocal lattice vectors and k is the wave vector inside the rst Brillouin zone. Each PW is augmented by an atomic-like function in every atomic sphere as described above. The solutions to the KohnSham equations are expanded in this combined basis set of LAPWs according to the linear variation method X wk cn /kn 6
n

and the coecients cn are determined by the RayleighRitz variational principle. The convergence of this basis set is controlled by a cuto parameter Rmt Kmax 69, where Rmt is the smallest atomic sphere radius in the unit cell and Kmax is the magnitude of the largest Kn vector in Eq. (6).

264

K. Schwarz, P. Blaha / Computational Materials Science 28 (2003) 259273

(APW, LAPW, APW + lo) mentioned above is given in [10], which contains many additional references. A combination of these schemes (LAPW and APW + lo) is the basis for the new WIEN2k program [11]. 3.2. The full-potential scheme The MTA was frequently used in the 1970s and works reasonable well in highly coordinated (metallic) systems such as face centered cubic (fcc) metals. However, for covalently bonded solids, open or layered structures, MTA is a poor approximation and leads to serious discrepancies with experiment. In all these cases a treatment without any shape approximation is essential. Both, the potential and charge density, are expanded into lattice harmonics (inside each atomic sphere) and as a Fourier series (in the interstitial region). P V rYLM ^ inside sphere; r V r PLM LM iKr outside sphere: K VK e 8 Thus their form is completely general so that such a scheme is termed full-potential calculation. The choice of sphere radii is not very critical in fullpotential calculations in contrast to MTA, in which one would obtain dierent radii as optimum choice depending on whether one looks at the potential (maximum between two adjacent atoms) or the charge density (minimum between two adjacent atoms). Therefore in MTA one must make a compromise but in full-potential calculations one can eciently handle this problem. 3.3. Relativistic eects In a solid that contains only light elements, a non-relativistic calculation is justied. As soon as heavier elements are involved, however, relativistic eects can no longer be neglected. Core states are treated fully relativistically. For valence states (and high lying semi-core states one shell below the valence states) relativistic eects of atoms in the medium range of atomic numbers (up to about 54) can be included in a scalar relativistic treatment. It

describes the main contraction or expansion of various orbitals (due to the Darwin s-shift or the massvelocity term) [12] but this scheme omits the spinorbit splitting. The latter can be included in a second-variational treatment [13]. An additional basis function in form of a local orbital can be added for p1=2 which diers from the scalar-relativistic p orbital [14]. For very heavy elements it may be necessary to solve Diracs equation, which has all these terms included. 3.4. Computational aspects In the newest version WIEN2k [11] the alternative basis set (APW + lo) is used inside the atomic spheres for those important -orbitals (partial waves) that are dicult to converge (outermost valence p-, d-, or f-states) or for atoms, where small atomic spheres must be used [9,10,15]. For all the other partial waves the LAPW scheme is used. In addition new algorithms for solving the computer intensive general eigenvalue problem were implemented. The combination of algorithmic developments and increased computer power has led to a signicant improvement in the possibilities to simulate relatively large systems on moderate computer hardware. Now PCs or a cluster of PCs can be eciently used instead of the powerful workstations or supercomputers that were needed about a decade ago. Several considerations are essential for a modern computer code and were made in the development of the new WIEN2k package [11]: Accuracy is extremely important in the present case. It is achieved by the well-balanced basis set, which contains numerical radial functions that are recalculated in each iteration cycle. Thus these functions adapt to eects due to charge transfer or hybridization, are accurate near the nucleus and satisfy the cusp condition. The PW convergence can essentially be controlled by one parameter, namely the cuto energy corresponding to the highest PW component. There is no dependence on selecting atomic orbitals or pseudo-potentials. It is a fullpotential and all electron method. Relativistic eects (including spin orbit coupling) can be

K. Schwarz, P. Blaha / Computational Materials Science 28 (2003) 259273

265

treated with a quality comparable to solving Diracs equation. Eciency and good performance should be as high a possible. The smaller matrix size of the new mixed basis APW + lo/LAPW helps to save computer time or allows studying larger systems. User friendliness is achieved by a web based graphical user interface (w2web) and is complemented by an automatic choice of default options. In addition an extensive Users Guide is provided. A modern software package needs a good graphical user interface (GUI). In WIEN2k this is im-

plemented as a web-based GUI, called w2web. As one example the structure generator is illustrated below and is discussed with Fig. 2 for the case of chromium dioxide CrO2 . One can give a title for the calculation, chromium dioxide, and select (the default) a relativistic treatment of the core states. In this example the known space group P42 /mnm, number 136, was selected from the list of 230 space groups. In this structure we have two inequivalent atoms. Next one species the lattice  parameters, either in A or in atomic units (a.u.). The latter are used in all the calculations but for  the input A is often more convenient. For CrO2 we have a tetragonal structure with a b but dierent from c. The three angles (a, b, c) are all

Fig. 2. Structure generator illustrated for CrO2 using the graphical user interface w2web.

266

K. Schwarz, P. Blaha / Computational Materials Science 28 (2003) 259273

90. Now the two inequivalent atoms must be specied. The rst is Cr, dened by the chemical symbol, from which the atomic number Z is generated automatically. The NPT (number of points in the logarithmic radial mesh) and r0 (the rst radial mesh-point) are set to default values, but the crucial parameter is the mun tin radius. It should be set as large as possible to save computer time but overlapping spheres must not occur. Then one must specify the atomic position, where only one of the equivalent atoms must be specied, whereas the others are generated using the symmetry operation of the specied space group. The procedure must be repeated for the second atom. This generates the most important le that characterizes the structure. In cases where the space group is not known, for example in an articially constructed super-cell, auxiliary programs can nd the space group. All other input les are generated during initialization with some default values, which need to be changed only in few special cases. On the left side of Fig. 2 one sees the various options, for example to initialize a calculation or to run a SCF cycle. There are several utility programs and for the most common tasks like band structure, electron density, or density of states, a guided menu simplies their generation. 3.5. WIEN2k hardware and software considerations The hardware and software are changing rapidly and require a continuous adaptation of an application software. We attempt to use high level languages for the program development to simplify such adaptations and try to hide the details from the regular user. In addition we provide installation support for most platforms which should make it easier for a users to optimally install the program package on the computer system that is available for a research group. For the interested reader we summarize some details and mention the most important aspects for running the WIEN2k code eciently. WIEN2k runs on any Unix/Linux platform from PCs, cluster, workstations to supercomputers.

FORTRAN90 is used mainly to allow dynamical storage allocation. The program size is automatically adapted to the problem size at runtime. The program consists of several separate programs (modules) that are linked by means of C-shell or perl-scripts. Architecture. The hardware in terms of processor speed, memory access, and communication is crucial. Depending on the given architecture, optimized algorithms and libraries are used during installation of the program package with support for most platforms. Portability requires the use of standards as much as possible, such as FORTRAN90, BLAS (level 3), MPI, SCALAPACK, etc., standard software that is essentially public domain. Besides that perl5, ghostview, gnuplot, a pdf-reader, Tcl/ Tk, and a web-browser should be available. Inversion symmetry determines whether the matrix elements are real or complex numbers. The latter case requires about twice as much memory and about four times more computer time for the calculations. The proper version of the relevant programs is automatically selected. A system (with inversion symmetry) requires about 128 MB (12 GB) for about 10 (100) atoms per unit cell. With modern processors not only the speed of the CPU is crucial but also the memory bus that determines how fast data can be accessed from memory. Parallelization. The program can run in parallel, either in a coarse grain version where each k-point is computed on a single processor, orif the memory requirement is larger than that available on one CPUin a ne grain scheme that requires special attention for the eigensolver, the most time consuming part. The k-point parallelization works very well on a cluster of PCs connected with a common NFS lesystem and needs only a slow network (100 MB/s). The ne-grain parallelization with MPI/SCALAPACK requires either a shared memory machine or a very fast communication. Depending on the system size both options, full (LAPACK) and iterative (Block Davidson method) diagonalization, are implemented to allow the selection of the most ecient routines.

K. Schwarz, P. Blaha / Computational Materials Science 28 (2003) 259273

267

4. Chromium dioxide as a representative application with WIEN2k The WIEN2k code can solve the KohnSham DFT equations with high accuracy. It provides many results like energy bands, electron density, total energy, forces acting on the atoms, spectra, and related quantities. Such results are often very useful to study various materials science problems. In this paper we select one system, for which we illustrate a few representative results, namely chromium dioxide, CrO2 , which has been intensively studied over the years. Originally the main interest was due to its use as a magnetic recording material. Then came the discovery that CrO2 was predicted to be a half-metallic ferromagnet [16], i.e. it has a metallic behavior in one spin direction, but a gap for the other spin. This property makes it very interesting for example in connection with spin-electronic devices, since there is a 100% spin polarization at the Fermi energy. CrO2 crystallizes in the rutile structure whose electronic structure was described in detail in [17]. The structure denition was explained in connection with the graphical user interface w2web in Section 3.4.

4.1. Energy band structure and density of states As a rst example we show the energy band structure of CrO2 , a half-metallic ferromagnet. In Fig. 3 the left two panels show the spin-up and the right the spin-down bands. The bands are only shown along a few high symmetry lines in the Brillouin zone. It can be clearly seen that spin-up electrons correspond to a metallic state, whereas the spin-down bands show a gap. This constitutes the half-metallic nature of CrO2 in agreement with previous work [16]. To each eigenvalue Enk we have the corresponding wave functions, from which we can calculate how much each atomic orbital contributes. This can be done by calculating the fraction of charge that is found in each atomic sphere or in the interstitial region. Inside each sphere the charge can be further decomposed into contributions from s-, p-, d-, f-, etc. partial waves. This is illustrated (Fig. 3) by showing the Cr-d and O-p contributions in the adjacent panels (a, b) for spin up and (c, d) for spin-down electrons. The low energy region ()2 to )7 eV) is dominated by O-p states with little admixture from Cr and only a small spin polarization. Thus these bands look very similar in both spin channels. For

(a)
5.0 4.0 3.0 2.0 1.0

(b)

(c)

(d)

Energy (eV)

0.0 -1.0 -2.0 -3.0 -4.0 -5.0 -6.0 -7.0 -8.0

EF

Fig. 3. Energy bands of ferromagnetic CrO2 for spin-up (a, b) and spin-down (c, d) electrons. The character of the bands is indicated by the circle size and shows the Cr-d admixture (a, c) and the O-p admixture (b, d), respectively.

268

K. Schwarz, P. Blaha / Computational Materials Science 28 (2003) 259273

the Cr-d bands, however, a large exchange splitting can be seen (compare the bands in panel a starting around )1 eV with the bands above +1 eV for the spin-down case in panel c). It should be noted that a classication according to irreducible representations is available and can be colorcoded. The corresponding density of states (DOS) is shown in Fig. 4. The charge decomposition of each state as discussed above for the character of the energy bands can also be used to dene the corresponding partial DOS showing the contributions from the chromium (dashed line) and oxygen (dotted line) sphere. In the range between )1.5 and )7.5 eV the O-DOS dominates with an admixture of Cr-DOS. The Cr-DOS dominates above )1.5 eV and shows a large exchange splitting that leads to the half-metallic character of CrO2 . For the spin-up electrons the Cr bands are signicantly

lower in energy and thus the Cr admixture in the oxygen bands is higher than in the spin-down case. It should be mentioned that the decomposition into partial DOS is model dependent and is affected by the choice of the sphere radii. If one would make the chromium sphere larger and the oxygen sphere smaller the details of the decomposition will slightly change. Therefore one should not put too much emphasis on quantitative details but focus on the trends. There is another aspect that should be understood, namely that the APW scheme uses a spatial decomposition of wave function and consequently its character changes from sphere to sphere in contrast to an LCAO scheme, where the atomic orbitals are atom-centered and extend way beyond the atomic region. For example, when an oxygen 2p orbital enters the neighboring chromium sphere this tail must be represented as Cr partial waves of a certain character inside the Cr sphere. We call this an osite component in contrast to on-site components like Cr-d or O-p states which reside mostly inside the Cr or O sphere, respectively. 4.2. Electron and spin density The electron density is a useful quantity for analysis and interpretation. It can be calculated for all occupied states, or a certain energy region of interest, or even for a single state Enk . This allows to visualize the character of bands and helps to interpret chemical bonding. In a magnetic case it is useful to visualize the spin magnetization density that is dened as the dierence between the spinup and spin-down electrons, respectively. Their integral gives the spin magnetic moment, which can be decomposed into atomic contributions corresponding to the atomic spheres. The sum of these spin densities q" and q# yields the total density. We illustrate the spin magnetization density for CrO2 in the plane through chromium and the two types of ligands, the equatorial and axial oxygen atoms (Fig. 5). A discussion about the comparison with experimental data and the interpretation is given in Section 5.1. The magnetization density around the chromium site shows a d-character with t2g symmetry forming bonds with the oxygen ligands. The

Fig. 4. Density of states (DOS) of ferromagnetic CrO2 for spinup (left) and spin-down (right) electrons. The partial DOS show the Cr and O contributions to the DOS.

K. Schwarz, P. Blaha / Computational Materials Science 28 (2003) 259273

269

Fig. 6. X-ray absorption (XAS) spectrum for the OK edge of CrO2 with a Gaussian broadening of 1 eV.

Fig. 5. Spin magnetization density, q" q# of ferromagnetic CrO2 in a plane through chromium and four of the six oxygen ligands. This gure is produced with XCrysDen [18] which provides a color coding for the positive and negative regions.

magnetization is localized near chromium and shows that even in an itinerant ferromagnet the magnetic moment comes from a local region in space. 4.3. Spectra Core-level spectra are particularly useful because their interpretation is simplied by the localized nature of the core hole, that leads to atom-selective information. A typical core wave function is completely conned inside the corresponding atomic sphere. Consequently the intensity of core-level spectra require only information from the atomic sphere, where the core hole occurs. These are for example X-ray emission and absorption spectra, or electron energy loss (XES, XAS, EELS). In XES a core hole is created and a transition occurs from the occupied valence states to the core hole governed by dipole selection rules. In XAS a transition from an occupied core state to an unoccupied valence state determines the intensity. Besides all broadening eects, such as lifetime, or spectrometric function the intensity of such spectra can be obtained from a ground state calculation.

We illustrate that for O-K XAS in Fig. 6 where a transition from the oxygen 1s level to unoccupied valence states that have O-p symmetry inside the oxygen sphere determine the spectrum. It consists mainly of the O-p DOS modied by proper transition matrix elements. In our case we can have transitions for spin-up or spin-down electrons that would lead to dierent spectra which will be averaged in the experiment. 4.4. Structure optimization In this example we illustrate two aspects, the calculation of a surface using a super-cell approach and the structural relaxations at the surface. A surface is modeled by a slab with a certain number of layers and a vacuum region separating these slabs. Such a unit cell is repeated with periodic boundary conditions in all three dimensions. It should be noted that such a slab has two surfaces which must be thick enough to reduce the articial interaction between the top and bottom surface. Furthermore, it is desirable that the central layer represents the bulk and the vacuum re gion is at least about 10 A. For CrO2 we model a (1 0 0) surface with Crtermination (see Fig. 7 (left panel)). In this case we have seven layers of chromium. One can see the distorted oxygen octahedra surrounding a Cr center. The initial structure is obtained by cutting

270

K. Schwarz, P. Blaha / Computational Materials Science 28 (2003) 259273

Fig. 7. Initial (unrelaxed) structure of a CrO2 slab with a Cr-terminated (1 0 0) surface (left panel). Structure of the CrO2 slab after relaxation (right panel). Chromium (large grey sphere) and oxygen (small black sphere).

such layers from a bulk material but keeping all structural parameters as in the bulk. Now we allow the structure to relax using the forces acting on the atoms to move them to the equilibrium positions, where the forces vanish and the corresponding total energy has a minimum. The result is shown in Fig. 7 (right panel) which shows that the structure remains approximately unchanged in the central layers, whereas a signicant relaxation takes place at the surface. The outermost Cr layer moves inward and the neighboring oxygens outward forming a buckled structure. Such a surface reconstruction has important consequences on the electronic structure and can be compared to experimental data.

5.1. Electron density The electron density is the key quantity in DFT and is determined in the self-consistent-eld (SCF) cycle. It is a ground state property that can be related to experiments. A Fourier transform of the electron density q leads to the static structure factor. In the experiment (e.g. X-ray diraction) additional factors must be considered, e.g. nite temperature, extinction, and absorption. A detailed comparison between theory and experiment was presented by Zuo et al. [19] for silicon, one of the best studied materials for which excellent samples exist. In contrast to the simple diamond structure of silicon with only two atoms per unit cell, the International Union for Crystallography (IUCr) has chosen corundum, Al2 O3 as the test compound for the multipole renement project [20]. In this work the renement of experimental data was investigated by using noise-free theoretical data for comparison. A detailed discussion of the comparison was given. As a third example we mention the polymorphs of Al2 SiO5 with the minerals andalusite, sillimanite and kyanite [21]. This structure contains four formula units (Z 4) and thus 32 atoms per unit

5. Other applications with WIEN2k In addition to the results explained in connection with CrO2 we provide a few examples that are important in our opinion, namely the electron density and its relation between theory and experiments, the electric eld gradients including interpretations, and a summary of features and topics for which references to the original literature is provided.

K. Schwarz, P. Blaha / Computational Materials Science 28 (2003) 259273

271

cell. In all three cases the structure consists of chains of octahedrally coordinated aluminum atoms parallel to the c-axis sharing edges with alternating silica tetrahedra and polyhedra of aluminum in 4-, 5-, and 6-fold coordination. These dierences had been analyzed on the basis of DFT calculations. The use of high-energy synchrotron radiation for studying the electron-density distribution is discussed for stishovite, SiO2 [22]. In addition to the comparison between theory and experiment a topological analysis is presented using the concept of atoms in molecules. Here the bond critical points, where rqr vanishes, were discussed. Another example is cuprite, for which the charge density and chemical bonding is discussed [23], a controversial issue also with respect to the high temperature superconductors. Since LDA has shortcomings for this system, dierent versions of LDA + U are tried to go beyond the conventional LDA or GGA. 5.2. Electric eld gradients Nuclear quadrupole interactions (NQI) can be measured by several experimental techniques such as Mssbauer (MB), nuclear magnetic resonance o (NMR), nuclear quadrupole resonance (NQR) or perturbed angular correlations (PAC). The isotopes of certain atoms have a nuclear quadrupole moment Q provided the nuclear quantum number I P 1. Representative examples are 57 Fe and 17 O. The electric eld gradient (EFG) is mainly determined by the anisotropy of the charge density surrounding the corresponding nucleus. It has been shown in 1985 that the EFG can be calculated from rst principles from the knowledge of the complete charge distribution [24]. An interpretation of EFGs was given for example for the four oxygen positions in the high temperature superconductor YBa2 Cu3 O7 [25], where it was shown that the dierent occupation numbers of the px , py , and pz orbitals inside the corresponding oxygen atomic spheres are directly related to the principle components of the EFG tensor. The dierence in occupation numbers can easily be understood on the basis of the electronic structure. In this structure two of these orbitals

point in a direction where no nearby neighbor is present and thus they are essentially non-bonding (i.e. fully occupied), whereas in the third direction (copper) neighbors are present leading to a large splitting between (Op Cud ) bonding and antibonding states, where the latter are only partially occupied due to the large splitting. This dierence in occupation numbers causes a non-spherical charge distribution that is responsible for the related EFG. It should be mentioned that the situation for the copper positions is dierent in two respects. On the one hand, at a Cu-position not only the p- but also the d-orbitals may contribute to the EFG often with opposite sign, which complicates the interpretation. On the other hand, copper oxides can be strongly correlated systems for which standard DFT may not be suciently accurate. This seems to be the case for Cu(2) in the buckled CuO layer, which shows large deviations between theory and experiment. The Cu(1) in the CuO linear chain has an EFG, however, that agrees well. Experiments always measure the NQI coupling, that is proportional to the product of Q, the nuclear quadrupole moment, and the EFG tensor. Often Q is known and can be used to convert the calculated EFG to the experimental coupling constant. In the case of Fe, the most important Mssbauer isotope, the value of Q was re-examo ined by calculating EFGs for a series of Fe compounds [26]. By relating these EFG values to the experimental data we could estimate the proportionality Q that was found twice as big as the previous value in literature. In the meantime the EFGs of many systems have been studied from metals to minerals [27,28]. In most cases the agreement between theory and experiment is often within 1020%. For strongly correlated systems larger deviations can be found [23,25]. 5.3. Further examples and new program features The WIEN code was used in many research topics to carry out electronic structure calculation to solve material science problems. A list of references, although by no means complete, can be found on the web-page (www.wien2k.at).

272

K. Schwarz, P. Blaha / Computational Materials Science 28 (2003) 259273

A few representative examples are summarized below. Test of new density functionals [29] LDA + U for highly correlated systems [30] Atomic core hole spectra [31] Optical properties [32] Detailed comparison with high resolution UPS spectroscopy for surface states, spin orbit splitting and spin splitting [3335] Unconventional heavy Fermion systems [36] Mixed valence systems [37] Intercalation compounds [38] Alkali-electro-sodalites [15] Thermoelectric materials [39]

is available. Therefore calculations allow us to predict a system to be an insulator, a metal or magnetic. The chemical bonding can be analyzed and this allows to make predictions, of how the system may change with deformations, under pressure or substitutions. Pressure is an easy parameter for DFT calculations in contrast to experiments. The opposite is true for estimating temperature eects which is easier for experiment than for theory. Pressure and nite temperature eects can be included using lattice vibrations, which may be included from phonon data or calculations. The examples given here are just a small selection to illustrate some possibilities that DFT and the WIEN2k code can provide.

At present several new features are developed and tested before they will be included in the WIEN2k package. These are the treatment of non-collinear magnetism (by R. Laskowski), non-linear optics (B. Olejnik), the calculation of phonons using a direct method (linking to the PHONON program by Parlinski [40]), and transport properties (G.K.H. Madsen). 6. Conclusion DFT calculations can provide useful information concerning the electronic structure of ordered crystal structures. The results depend on the choice of the exchange and correlation potential that is used in the calculations. In comparison with real systems it should be stressed that all kinds of imperfections are ignored. Some of them can be simulated. For example, it is possible to represent defects, e.g. an impurity, by studying a super-cell that consists (in a simple case) of 2 2 2 times the original unit cell. When one atoms in this large unit cell is replaced by an impurity atom, the neighboring atoms can relax to new equilibrium positions. Such a super-cell is repeated periodically and approximately represents an isolated impurity. This example illustrates how theory can carry out computer experiments. Hypothetical or articial structures can be considered and their properties, irrespective of whether or not they can be made, can be calculated. Complete information concerning the electronic, optic or magnetic properties

Acknowledgements The Austrian Science Fund (FWF Projects P8176, P9213, M134, and SFB 011/8) supported this work at the TU Wien. It is a pleasure to thank all coworkers who contributed to the present paper: G.K.H. Madsen, who implemented the new APW + lo scheme, J. Luitz (graphical user interface) and D. Kvasnicka (matrix diagonalization). The large community of WIEN users has contributed many improvements of the code (see www.wien2k.at). Their contributions are acknowledged with great thanks. In particular we want to thank the following persons and their group who substantially improved the features of the WIEN code. C. Ambrosch-Draxl (Graz) optics L. Nordstrm (Uppsala) APW + local orbitals o D. J. Singh (Naval Research Lab, Washington, DC) APW + local orbitals M. Scheer (FHI, Berlin) parallelization, optimization, surfaces P. Novak (Prague) relativistic eects (spin orbit), LDA + U V. Petricek (Prague) space groups C. Persson (Uppsala) irreducible representations B. Yanchitsky, A. Timoshevskii (Kiev) symmetry

K. Schwarz, P. Blaha / Computational Materials Science 28 (2003) 259273

273

J. Sofo (Penn State) topological analysis, Baders analysis R. Laskowski (Vienna) non-collinear magnetism B. Olejnik (Vienna) non-linear optics References
[1] P. Hohenberg, W. Kohn, Phys. Rev. B 136 (1964) 864. [2] W. Kohn, L.S. Sham, Phys. Rev. A 140 (1965) 1133. [3] J.P. Perdew, K. Burke, M. Ernzerhof, Phys. Rev. Lett. 77 (1996) 3865. [4] O.K. Andersen, Phys. Rev. B 12 (1975) 3060. [5] D.J. Singh, Plane Waves, Peudopotential and the LAPW Method, Kluwer Academic Publishers, Boston, Dortrecht, London, 1994. [6] P. Blaha, K. Schwarz, P. Sorantin, S.B. Trickey, Comput. Phys. Commun. 59 (1990) 399415. [7] D.J. Singh, Phys. Rev. B 43 (1991) 6388. [8] E. Sj stedt, L. Nordstrm, D.J. Singh, Solid State Como o mun. 114 (2000) 15. [9] G.H.K. Madsen, P. Blaha, K. Schwarz, E. Sjstedt, L. o Nordstrm, Phys. Rev. B 64 (2001) 195134-1. o [10] K. Schwarz, P. Blaha, G.K.H. Madsen, Comput. Phys. Commun. 147 (2002) 71. [11] P. Blaha, K. Schwarz, G.K.H. Madsen, D. Kvasnicka, J. Luitz, An augmented plane wave plus local orbitals program for calculating crystal properties, Vienna University of Technology, Austria, 2001, ISBN 3-9501031-1-2. [12] D.D. Koelling, B.N. Harmon, Solid State Phys. 10 (1977) 3107. [13] A.H. MacDonnald, W.E. Picket, D.D. Koelling, J. Phys. C: Solid State Phys. 13 (1980) 2675. [14] J. Kunes, P. Novk, R. Schmid, P. Blaha, K. Schwarz, a Phys. Rev. B 64 (2001) 153102. [15] G.H.K. Madsen, B.B. Iversen, P. Blaha, K. Schwarz, Phys. Rev. B 64 (2001) 195102-1-6. [16] K. Schwarz, J. Phys. F: Met. Phys. 16 (1986) L211. [17] P.I. Sorantin, K. Schwarz, Inorg. Chem. 31 (1992) 567. [18] A. Kokalj, J. Mol. Graph. Model. 17 (1999) 176. [19] J.M. Zuo, P. Blaha, K. Schwarz, J. Condens. Matter 9 (1997) 7541.

[20] S. Pillet, M. Souhassou, C. Lecomte, K. Schwarz, P. Blaha, M. Rerat, L. Lichanot, P. Roversi, Acta Cryst. A 57 (2001) 290. [21] M. Iglesias, K. Schwarz, P. Blaha, D. Baldomir, Phys. Chem. Minerals 28 (2001) 67. [22] A. Kirfel, H.-G. Krane, P. Blaha, K. Schwarz, T. Lippmann, Acta Cryst. A 57 (2001) 663. [23] R. Laskowski, P. Blaha, K. Schwarz, Phys. Rev. B 67 (2003) 075102. [24] P. Blaha, K. Schwarz, P. Herzig, Phys. Rev. Lett. 54 (1985) 1192. [25] K. Schwarz, C. Ambrosch-Draxl, P. Blaha, Phys. Rev. B 42 (1990) 2051. [26] P. Dufek, P. Blaha, K. Schwarz, Phys. Rev. Lett. 75 (1995) 3545. [27] P. Blaha, K. Schwarz, W. Faber, J. Luitz, Hyperne Int. 126 (2000) 389. [28] B. Winkler, P. Blaha, K. Schwarz, Am. Minerolog. 81 (1996) 545. [29] J.P. Perdew, S. Kurth, J. Zupan, P. Blaha, Phys. Rev. Lett. 82 (1999) 2544. [30] P. Mohn, C. Persson, P. Blaha, K. Schwarz, P. Novk, H. a Eschrig, Phys. Rev. Lett. 87 (2001) 196401-1. [31] J. Luitz, M. Maier, C. Hebert, P. Schattschneider, P. Blaha, K. Schwarz, B. Jourey, Eur. Phys. J. B 21 (2001) 363. [32] C. Ambrosch-Draxl, J. Majewski, P. Vogl, G. Leising, Phys. Rev. B 51 (1995) 9668. [33] J. Schfer, E. Rotenberg, G. Meigs, S.D. Kevan, P. Blaha, a S. Hfner, Phys. Rev. Lett. 83 (1999) 2069. u [34] J. Schfer, E. Rotenberg, S.D. Kevan, P. Blaha, R. a Claessen, R.E. Thorne, Phys. Rev. Lett. 87 (2001) 196403-1. [35] G. Nicolay, F. Reinert, S. Hfner, P. Blaha, Phys. Rev. B u 65 (2001) 033407-1. [36] D.J. Singh, P. Blaha, K. Schwarz, I.I. Martin, Phys. Rev. B 60 (1999) 1635916363. [37] P. Novk, F. Boucher, P. Gressier, P. Blaha, K. Schwarz, a Phys. Rev. B 63 (2001) 235114-1. [38] X. Rocquefelte, F. Boucher, P. Gressier, G. Ouvrard, P. Blaha, K. Schwarz, Phys. Rev. B 62 (2000) 2397. [39] G.K.H. Madsen, K. Schwarz, P. Blaha, D.J. Singh, Phys. Rev. B, in press (2003). [40] K. Parlinski, Software PHONON, 2002.