An Introduction to Density Functional Theory

1 An Introduction to Density Functional Theory N. M. Harrison Department of Chemistry, Imperial College of Science Technology and Medicine, SW7 2AY, London and CLRC, Daresbury Laboratory, Daresbury, Warrington, WA4 4AD. For the past 30 years Density Functional Theory has been the dominant method for the quantum mechanical simulation of periodic systems. In recent years it has also been adopted by quantum chemists and is now very widely used for the simulation of energy surfaces in molecules. In this lecture we introduce the basic concepts underlying Density Functional Theory and outline the features that have lead to its wide spread adoption. Recent developments in exchange correlation functionals are introduced and the performance of families of functionals reviewed.

2 The lecture is intended for a researcher with little or no experience of quantum mechanical simulations but with a basic (undergraduate) knowledge of quantum mechanics. We hope to provide sufficient background to enable informed judgements on the applicability of a particular implementation of Density Functional Theory to a specific problem in materials simulation . For those who wish to go more deeply into the formalism of Density Functional Theory there are a number of reviews and books aimed at intermediate and advanced levels available in the literature [1,2,3]. Where appropriate source articles are referred to in the text. 1. The Solution of the Schr dinger Equation During the course of this lecture we will be primarily concerned with the calculation of the ground state energy of a collection of atoms.

3 The energy may be computed by solution of the Schr dinger equation which, in the time independent, non- relativistic, Born-Oppenheimer approximation is 1 ;.. H (r1 , r2 ,.., rN ) = E (r1 , r2 ,.., rN ). Equation 1. 1. Atomic units are used throughout. 1. The Hamiltonian operator, H, consists of a sum of three terms; the kinetic energy, the interaction with the external potential (Vext ) and the electron-electron interaction (Vee). That is;.. 1 N N.. 1. H = + +. 2. V ext i < j | ri r j |. i 2 i Equation 2. In materials simulation the external potential of interest is simply the interaction of the electrons with the atomic nuclei;. Nat Z . V ext = . | ri R |. Equation 3. Here, ri is the coordinate of electron i and the charge on the nucleus at R is Z . Note that in order to simplify the notation and to focus the discussion on the main features of DFT the spin coordinate is omitted here and throughout this article.

4 Equation 1 is solved for a set of subject to the constraint that that the are anti- symmetric they change sign if the coordinates of any two electrons are interchanged. The lowest energy eigenvalue, E0 , is the ground state energy and the probability Density of finding an electron with any particular set of coordinates {ri} is | 0 |2 . The average total energy for a state specified by a particular , not necessarily one of the eigenfunctions of Equation 1, is the expectation value of H, that is;.. E[ ] = * H dr H . Equation 4. The notation [ ] emphasises the fact that the energy is a Functional of the wavefunction. The energy is higher than that of the ground state unless corresponds to 0 which is the variational theorem;. E[ ] E0. Equation 5. 2. The ground state wavefunction and energy may be found by searching all possible wavefunctions for the one that minimises the total energy.

5 Hartree-Fock Theory consists of an ansatz for the structure of - it is assumed to be an antisymmetric product of functions ( i) each of which depends in the coordinates of a single electron, that is;. HF =. 1. N! [. det 1 2 3 .. N ]. Equation 6. where, det indicates a matrix determinant [4]. Substitution of this ansatz for . into the Schr dinger equation results in an expression for the Hartree Fock energy;. 1 N 2 . E HF = i* (r ) i + Vext i (r )dr 2 i . 1 N (r ) (r ) (r ) (r ). * *. + i 1 i 1 j 2 j 2 dr1dr2. 2 i,j ri r j 1 N i ( r1 ) j ( r1 ) i ( r2 ) j (r2 ). * *. dr1dr2. 2 i, j ri r j Equation 7. The second term is simply the classical Coulomb energy written in terms of the orbitals and the third term is the exchange energy. The ground state orbitals are determined by applying the variation theorem to this energy expression under the constraint that the orbitals are orthonormal.

6 This leads to the Hartree-Fock (or SCF). equations;. 1 2 (r ' ) . + v ext ( r) + dr' i ( r) + v X (r , r ' ) i ( r' )dr ' = i i (r ). 2 r r ' . Equation 8. Where the non-local exchange potential, vX, is such that: 3. N j (r ) *j (r ' ). v X (r, r ' ) i (r ' )dr ' = i ( r' ) dr '. j r r'. Equation 9. The Hartree-Fock equations describe non-interacting electrons under the influence of a mean field potential consisting of the classical Coulomb potential and a non-local exchange potential. From this starting point better approximations (correlated methods) for and E0 are readily obtained but the computational cost of such improvements is very high and scales prohibitively quickly with the number of electrons treated (for an excellent Introduction see ref. [4]). In addition, accurate solutions require a very flexible description of the wavefunction's spatial variation, a large and basis set is required which also adds to the expense for practical calculations.

7 Many correlated methods have been developed for molecular calculations [4]. The cost of the most commonly used methods, MP2, MP3, MP4, CISD, CCSD, CCSD(T) formally scales with the number of electrons raised to the power of 5,6,7,6,6,7 respectively. In most cases CCSD(T) calculations are of sufficient accuracy to determine the chemical properties of systems to sufficient accuracy to predict chemical properties (stability, reaction rates ) however, due to the computational expense the routine application of such methods to realistic models of systems of interest is not practical and not likely to become so despite rapid advances in computer technology. In this context we mention recent advances in the solution of the Schr dinger equation using variational quantum Monte Carlo approach [5].

8 The discussion above has established that direct solution of the Schr dinger equation is not currently feasible for systems of interest in condensed matter science . this is a major motivation for the development and use of Density Functional Theory . The question that arises is Is it necessary to solve the Schr dinger equation and determine the 3N dimensional wavefunction in order to compute the ground state energy ? 2. Avoiding the Solution of the Schr dinger Equation The Hamiltonian operator (Equation 2) consists of single electron and bi- electronic interactions operators that involve on the coordinates of one or two electrons only. In order to compute the total energy we do not need to know the 3N. dimensional wavefunction. Knowledge of the two-particle probability Density that is, the probability of finding an electron at r1 and an electron at r2 is sufficient.

9 A quantity of great use in analysing the energy expression is the second order Density matrix, which is defined as: 4. N ( N 1). P2 (r1' , r2' ; r1 , r2 ) =. 2 * (r1' , r1' ,.., rN' ) (r1 , r2 ,.., rN )dr3dr4 ..drN. Equation 10. The diagonal elements of P2 , often referred to as the two-particle Density matrix or pair Density , are;. P2 (r1 , r2 ) = P2 (r1 , r2 ; r1 , r2 ). This is the required two electron probability function and completely determines all two particle operators. The first order Density matrix is defined in a similar manner and may be written in terms of P2 as;. 2. P1 (r1' ; r1 ) =. N 1 P2 (r1' , r2 ; r1 , r2 ) dr2. Equation 11. Given P1 and P2 the total energy is determined exactly;. 1 2 N a t Z 1. E = tr H P = 1 P1 (r1' , r1 ) dr1 + P2 (r1 , r2 ) dr1dr2.

10 2 | r1 R | r1 = r1' | r1 r2 |. Equation 12. We conclude that the diagonal elements of the first and second order Density matrices completely determine the total energy. This appears to vastly simplify the task in hand. The solution of the full Schr dinger equation for is not required it is sufficient to determine P1 and P2 - and the problem in a space of 3N coordinates has been reduced to a problem in a 6 dimensional space. Approaches based on the direct minimisation of E(P1 ,P2 ) suffer from the specific problem of ensuring that the Density matrices are legal that is, they must be constructible from an antisymmetric . Imposing this constraint is non trivial and is currently an unsolved problem [6,7]. In view of this we conclude that Equation 12. does not lead immediately to a reliable method for computing the total energy without calculating the many body wavefunction.