Comparison of exchange-correlation functionals: from LDA to GGA and beyond

Martin Fuchs Fritz-Haber-Institut der MPG, Berlin, Germany

Density-Functional Theory Calculations for Modeling Materials and Bio-Molecular Properties and Functions - A Hands-On Computer Course, 30 October - 5 November 2005, IPAM, UCLA, Los Angeles, USA

Density-functional theory
DFT is an exact theory of the ground state of an interacting many-particle system:

E0 = min Ev [n]
nN

E0 = min |Hv |
N

by Hohenberg-Kohn theorem: v(r) v[n(r)]

Total energy density functional (electrons: w(r, r ) = 1/|r r |) Z Ev [n] = min |T + W | + n(r)v(r)d = F [n]
n

n(r)v(r) d

Non-interacting case: w(r, r ) 0

F [n] min |T | = [n]|T |[n] = Ts[n]

n

[n] is a Slater determinant

Interacting case rewritten:

1 Ev [n] = Ts[n] + Exc [n] + 2

n(r)w(r, r )n(r )d d +

n(r)v(r) d

XC energy functional: kinetic correlation energy & exchange and Coulomb correlation energy

Exc [n] =: [n]|T |[n] Ts[n] + [n]|W |[n]

1 2

n(r)w(r, r )n(r )d d

Kohn-Sham scheme
Minimization of Ev [n] with n(r) N determines groundstate:

Ev [n] = n(r)

Ts[n] F [n] KS + v(r) = + v ([n]; r) = n(r) n(r)

. . . eective non-interacting system Kohn-Sham independent-particle equations:

2 2

+ v KS (r) i(r) = P
i

i i (r) with v (r) =

KS

w(r, r )n(r )d +

Exc [n] + v(r) , n(r)

density n(r) = occupancies

i

fi|i(r)|2,
i

< : fi = 1,

= : 0 fi 1,

> : fi = 0 (aufbau principle).

i

non-interacting kinetic energy treated exactly Ts[n] = classical electrostatics (e-n & e-e) treated exactly only need to approximate Exc [n] and vxc ([n]; r) =

R P

fi| i(r)|2d
determines accuracy in practice!

Exc [n] n(r)

applied is spin density functional theory: n, treated as separate variables

beware: v KS (r) is a local operator. Direct approximation of whole F [n] possible too: DFT with beware: generalized Kohn-Sham schemes1 and nonlocal eective potentials, e.g. Hartree-Fock eqs.
1

Seidl, Grling, Vogl, Majewski, Levy, Phys Rev B 53, 3764 (1996). o

Local-density approximation
LDA Exc [n]

n(r)exc (n(r))d local dependence on density or Wigner-Seitz radius rs = ( 4 n)1/3 3 XC energy per electron is that of the homogeneous electron gas, exc [n] = ehom (n)|n=n(r) xc  exchange part ehom (n) known analytically x  correlation part ehom (n) known c analytically for rs 0 and rs numerically exact for 2 < rs < 100 from QMC data1 . . . as parametrization interpolating over all rs PW91 Perdew-Zunger VWN2 , , workhorse in DFT applications to solids real systems are far from jellium-like homogeneity . . . why does LDA work at all? . . . where does it fail? . . . how to improve beyond it?

hom

0.0

exc (Ry)

10.0

20.0

30.0

0.1

1.0

10.0

100.0

rs (bohr)
1 Ceperley, Alder, Phys Rev Lett 45, 566 (1980). 2

Perdew, Wang, Phys Rev B 45, 13244 (1992); Perdew, Zunger (1980); Vosko, Wilk, Nussair (1980).

Performance of the LDA

structural, elastic, and vibrational properties often good enough  crystal bulk lattice constants accurate to within 3%, usually underestimated  bulk moduli somewhat too large, > 10% error not uncommon for d-metals  phonons somewhat too sti binding energies are too negative (overbinding), up to several eV # cohesive energies of solids, but formation enthalpies often o.k. # molecular atomization energies, mean error (148 molecules G2 set) 3 eV activation energies in chemical reactions unreliable # too small/absent, e.g. for H2 on various surfaces (Al, Cu, Si, ...) relative stability of crystal bulk phases can be uncertain # SiO2 high pressure phase more stable than zero pressure phase # underestimated transition pressure e.g. for diamond -tin phase transitions in Si & Ge # magnetic phases electronic structure can be usefully interpreted (density of states, band structures), except for band gaps (a more fundamental issue than LDA!)

View on XC through the XC hole

Denition of the XC energy by a coupling constant integration of the e-e interaction 1:

E([n]; ) = [n]|T + V + W |[n] = min ...

n

by Hohenberg-Kohn theorem: n(r) v([n]; r) for any 0 1 = 0 non-interacting case/ Kohn-Sham potential, v=0([n]; r) = v KS ([n]; r) = 1 external potential, v=1([n]; r) = v(r) Z 1 d E([n]; ) groundstate energy: E([n]; 1) = E([n]; 0) + d d 0
Exc [n] =

R1
0

[n]|W |[n] d Z w(r, r )n(r)n(r )

1 2

n(r)w(r, r )n(r )d d P
p,q

1 X, XX = 2

[n]|

(r) (r )q (r )p(r)|[n] p q n(r)n(r )

d d

 XC in terms of the pair correlation function

PSfrag replacements g([n]; r, r ) r

 XC in terms of the adiabatic connection integrand (scaled e-e potential energy)

g(r, r )

may distinguish exchange and correlation, e.g. Exc = Ex + Ec

1

cf. DFT books by Dreizler and Gross & Yang and Parr

. . . coupling constant averaged XC hole

-integration implies coupling constant averaged pair correlation function g ([n]; r, r ) =
Identify the XC energy as

g([n]; r, r )d
0

1 Exc [n] = 2

d n(r)

d n(r ){([n]; r, r ) 1} w(r, r ) g

interpretation: the electron density n(r) interacts with the electron density of the XC hole
Pauli exclusion principle & Coulomb repulsion
XC hole (n)

0.0

+ correlation

nxc ([n]; r, r ) = n(r ){([n]; r, r ) 1} g

exchange

local density approximation corresponds to nxc ([n]; r, r ) = n(r){ g

hom

0.5

homogeneous electron gas

(n(r); |r r |) 1}
0 1 2 3 4 5 6

PSfrag replacements always centered at reference electron & spherical

k F |r r |
(NB: correlation part varies with kF )

From the XC hole to the XC energy

Focus on Exc [n] =

n(r)exc ([n]; r)d (... component of the total energy) w =

1 |rr |

1 |u|

for exc only the angle averaged XC hole matters (and LDA hole is always spherical) nxc (r, r + u) = X
lm

nlm(r, u)Ylm(u)

xc

1 Exc [n] = 2

n(r)

nxc (r, r + u) 1 du d u 2

n(r)

nxc (r, u) 2 00 u du d u

for Exc only the system & angle averaged XC hole matters Z 1 xc n(r)n00(r, u)d nxc (u) = N
XC energy in terms of averaged XC hole1 Z N nxc (u) 2 Exc [n] = u du 2 0 u

N average XC energy per electron

LDA & GGA approximate average holes rather closely work mostly o.k.
1

Perdew et al., J Chem Phys 108, 1552 (1998).

Sum rule and other constraints on the XC hole

sum rule (constrains global behavior) Z nxc ([n]; r, r )dr = 1 ... average hole too
 LDA:

0 average XC hole

approximate

n(r)nhom (n(r), r, r )dr = 1 xc

sum rule

exact

on-top hole nxc (0) (xes value at u = 0)  LDA: accurate (exact in some limits) for correlation, exact for exchange cusp condition (constrains behavior around u = 0) nxc (u) = nxc (0) + n(0) u u=0

cusp ontop

2 3 separation u

 LDA: correct

LDA works well because the LDA (average) XC hole is that of a physical system, jellium
Beware: pointwise behavior of nLDA(r, r ) may be incorrect (e.g. outside metal surface), xc system averaging unweights tail and near-nucleus regions XC potential vxc ([n]; r) for LDA can be locally poor, but again less so for the system average

Generalized Gradient Approximation for Exc

Gradient expansion of XC energy and (later) hole: nxc ([n]; r, u) nxc (r, u = 0) + and generalized to imposing constraints to meet R  e.g. nx (r, u) < 0, nx (r, u)dr = 1, by real space cutos numerical GGA  scaling relations and bounds on Exc by analytic approximation to numerical GGA parameter-free GGA by Perdew-Wang PW91 simplied in PBE GGA  contains LDA & retains all its good features Earlier: analytic model or ansatz + empirical parameter(s) Langreth-Mehl (C), Becke 86(X) + Lee-Yang-Parr(C) BLYP, ... Alternative: analytic ansatz + (many) tted parameters (e.g. t to thermochemical data)  see accuracy limit of GGA functionals, can be better & worse than PBE Generic GGA XC functional Z GGA LDA GGA Exc [n] = n ex (n) Fxc (n, s)
u nxc |u=0

+ ...,

n=n(r) d,

| n| s= 2kF n n=n(r)

Enhancement factor Fxc (n, s) over LDA exchange: function of density and scaled gradient understanding how GGAs work Calculations with GGAs are not more involved than with LDA, except that vxc [n; r] = vxc (n, in, i j n)|n(r)

Dierences in present GGAs

PBE
1.8
rs=100

Zhang et al. revised revPBE

1.8
rs=100

1.8
rs=100

BLYP
50

enhancement factor

enhancement factor

1.6

50

enhancement factor

1.6

50

1.6

PBE XC

1.4

10 5

PSfrag replacements
r =0

1.4

10 5

PSfrag replacements
Zhang

1.4

10 5

ements

1.2

s enhancement factor scaled gradient

1.2

enhancement factor scaled gradient

3

1.2

BLYP

1.0

1.0

1.0

scaled gradient

scaled gradient

scaled gradient

... WC-PBE Wu & Cohen 05 xPBE Xu & Goddard 04 HCTH Handy et al. 01 RPBE Hammer et al. 99 revPBE Zhang et al. 98 PBE GGA Perdew et al. 96 PW91 Perdew & Wang 91 PW91 Perdew & Wang 91 BP Becke & Perdew 88 BLYP Becke & Lee et al. 88 ...

revPBE & BLYP more nonlocal than PBE GGA molecules: more accurate atomization energies solids: bondlengths too large, cohesive energies too small LYP correlation incorrect for jellium more local GGAs will make lattice constants smaller (e.g. Tinte et al. PRB 58, 11959 (1998)), but make binding energies more negative

... how GGAs change total energies

Analysis in terms of selfconsistent LDA density shows how GGAs work:

Etot [n

GGA

GGA

] = Etot [n

LDA

LDA

] + Exc [n

GGA

LDA

] Exc [n

LDA

LDA

] + O(n

GGA

LDA 2

Spectral decomposition in terms of s: ZZ GGA LDA Exc EXC =

s=0

GGA exc (r)

LDA exc (r)

(s s(r))d ds

0.05
energy per atom (eV)

energy per atom (a.u.)

correlation

Al fcc

0.0

PBE
differential

X+C

0.00
PBE rev PBE exchange

0.2

integrated

-0.3 (0.5) eV
revPBE 0.5 0
fcc Al

0.05 0 1

PSfrag replacements
4

-0.5 (0.7) eV
1 2 3 scaled gradient s 4

2 3 scaled gradient s

only 0 s 4 contribute, similar analysis can be made for n(r) for more see e.g. Zupan et al., PRB 58, 11266 (1998)

Cohesive properties in GGA

Bulk lattice constants GGA increase due to more repulsive core-valence XC.
s l C l C lAs aA u e a a N N A C Si G Si A G C

Cohesive energies valence eect.

GGA reduction mostly

s l C l C lAs aA u e a a N N A C Si G Si A G C

5
PW91GGA

1.5
expt.

LDA inconsistent GGA

error (eV)

error (%)

1.0 0.5 0.0 0.5

PW91GGA expt.

inconsistent GGA LDA

for comparison of LDA, GGA, and Meta-GGA see Staroverov, Scuseria, Perdew, PRB 69, 075102 (2004)

... energy barriers: H2 on Cu(111)

barrier to dissociative adsorption - PW91-GGA 0.7 eV - LDA <0.1 eV total energy of free H2 - PW91-GGA 31.8 eV ( expt.) 30.9 eV - LDA

energy TS
ads Ea ??

Cu + H2: LDA-GGA 1 eV
des Ea

reaction coordinate Hammer et al. PRL 73, 1400 (1994)

Phase transition -tin - diamond in Si

Spectral decomposition in terms of n

2.0

Relative volume / rs

1.5

-tin

diamond-structure

1.0

c a -tin structure a

0.5

diamond

0.0

Density parameter rs [bohr]

diamond: a0 = 10.26 bohr, semiconductor

-tin:

c/a = 0.552, metallic

Phase transition -tin - diamond in Si

Gibbs construction:
dia Et + ptVt = Et + ptVtdia ,

pexp = 10.3 . . . 12.5 GPa t

25
0.6

tin

diamond

transition pressure (GPa)

BLYP

energy (eV/atom)

20

0.4
GGA LDA

15
expt

BP PW91 revPBE PBE

0.2

10
LDA

0.0 0.6 0.7 0.8 0.9 1.0 1.1

0.2

relative volume V/Vexp

GGA increases transition pressure, inhomogeneity eect use of LDA-pseudopotentials insucient

0.3 0.4 0.5 0.6 energy change (eV/atom)

Moll et al, PRB 52, 2550 (1995); DalCorso et al, PRB 53, 1180 (1996); McMahon et al, PRB 47, 8337 (1993).

Performance of PBE GGA vs. LDA

 atomic & molecular total energies are improved  GGA corrects the LDA overbinding: average error for G2-1 set of molecules: +6 eV HF 1.5 eV LSDA 0.5 eV PBE-GGA 0.2 eV PBE0 hybrid 0.05 eV goal, better cohesive energies of solids  improved activation energy barriers in chemical reactions (but still too low)  improved description of relative stability of bulk phases  more realistic for magnetic solids  useful for electrostatic hydrogen bonds # Van der Waals (dispersion) forces not included! # GGA softens the bonds increasing lattice constants decreasing bulk moduli no consistent improvement

LDA yields good relative bond energies for highly coordinated atoms, e.g. surface energies, diusion barriers on surfaces GGA favors lower coordination (larger gradient!), not always enough where LDA has a problem, e.g. CO adsorption sites on transition metal surfaces signicance of GGA ? # GGA workfunctions for several metals turn out somewhat smaller than in LDA one-particle energies/bands close to LDA

Comparing LDA and (dierent) GGAs gives an idea about possible errors!

Beyond GGA: orbital dependent XC functionals

Kohn-Sham non-interacting system make i[n] density-functionals: n v KS [n; r] i

LDA and GGA are explicit density-functionals Exc [n] Implicit density-functionals formulated in terms of i[n]? more exible for further improvements self-interaction free: Exc [n]|N =1 = 0, vxc (|r| ) = 1 , r nonspherical nxc (r, r ) ... ?
Start with exact exchange :

1 Exchange-energy Ex [{i}] = 2

ZP

i,j

i(r)j (r )i(r )j (r) |r r |

d d
KS and HF orbitals dierent!

v same as in Hartree-Fock, but

2 2

v KS (r)|i(r) =

i i (r)

Groundstate ... in KS-DFT: optimized eective potential OEP method, local KS potential in Hartree-Fock: variation with respect to orbitals yield HF eqs., nonlocal eective potential HF KS E0 E0 EXX

Hybrid functional: mixing exact exchange with LDA/GGA

XC = EXX + LDA correlation is less accurate (underbinds) than XC = LDA for molecules challenge: correlation functional that is compatible with exact exchange? hybrid functionals to interpolate adiabatic connection, Uxc [n] Z 1 Exc [n] = Uxc [n]() d Uxc [n]() = [n]|W |[n] UHartree [n]
0

molecular dissociation: Uxc = molecules - atoms ...

looks like

Uxc ()

Ex
LDA,GGA too negative

coupling strength

Ec Tc
LDA, GGA good

=0 = KS KS system

=1 physical system

. . . hybrid functionals
How they are dened: How hydbrids and GGAs work: Adiabatic connection N2 2N
2 0

use exact exchange for = 0 and a local functional for = 1

Hybrid functional = interpolation n o hyb GGA GGA Exc = Exc + a Ex Ex

exchange only "exact"

(eV)
-2

Uxc()

mixing parameter a = 0.16 . . . 0.3 from tting thermochemical data a = 1/4 by 4th order perturbation theory PBE0 = PBE1PBE

PBE0 hybrid
-4

PBE GGA
-6 0 0.5 1

B3LYP 3-parameter combination of Becke X-GGA, LYP C-GGA, and LDA molecular dissociation energies on average within 3 kcal/mol 0.1 eV (but 6 times larger errors can happen)

coupling strength

XC contributions to binding: (eV) PBE PBE0 exact

Eb -10.58 -9.62 -9.75

Ex -1.82 -1.01 1.39

Ec -2.49 -2.49 -5.07

... X and C errors tend to cancel!

Becke, J Chem Phys 98, 5648 (1993); Perdew et al, J Chem Phys 105, 9982 (1996). comprehensive comparison: Staroverov, Scuseria, Tao, Perdew, J Chem Phys 119, 12129 (2003)

XC revisited: role of error cancellation between X and C

... or: when GGAs or hybrids are not enough

see e.g. Baerends and Gritsenko, JCP 123, 062202 (2005)

... orbital dependent functionals

Meta-GGAs XC energy functional

MGGA Exc [n]

MGGA n(r)exc (n,

kinetic energy density of non-interacting electrons

ts(r) =

1 2

Pocc
i

n, ts)

n(r)

fi| i(r)|2

MGGA 1 Exc [i] OEP or HF style treatment of = ... i(r) 2

exc ts

TPSS: Tao, Perdew, Stavroverov, Scuseria, Phys Rev Lett 91, 146401 (2003): + vx nite
PKZB: Perdew, Kurth, Zupan, Blaha, Phys Rev Lett 82, 2544 (1999): XC non-empirial LDA limit , , Van Voorhis, Scuseria, J Chem Phys 109, 400 (1998): XC highly tted no LDA limit Colle, Salvetti, Theoret Chim Acta 53, 55 (1979): C, no LDA limit BLYP GGA Becke, J Chem Phys 109, 2092 (1998): XC + exact exchange tted

TPSS accomplishes a consistent improvement over (PBE) GGA PKZB improved molecular binding energy, but worsened bond lengths in molecules & solids hybrid functionals on average still more accurate for molecular binding energies TPSS provides sound, nonempirical basis for new hybrids TPSSh next step: MGGA correlation compatible with exact exchange?

ACFDT XC: including unoccupied Kohn-Sham states

Adiabatic connection: KS system = 0 physical system = 1 R1 |Wee | d U [n] Exc [n] =
0

Fluctuation-dissipation theorem:

Wxc () =: |Wee |

1 = 2

Z e2 (iu, r, r ) du n(r)(r r ) drdr |r r | 0

. . . using dynamical density response

From noninteracting Kohn-Sham to interacting response by TD-DFT P (r)j (r)(r )i(r ) . . . KS eigenvalues i [n] i j 0(iu, r, r ) = 2 i,j i u( j i ) i([n], r) . . . KS orbitals
hxc (iu) = 0(iu) + 0(iu) K (iu) (iu)

Dyson equation 6dim.

. . . using Coulomb and XC kernel from TD-DFT In principle ACFDT formula gives exact XC functional In practice starting point for fully nonlocal approximations
hxc RPA: K = |r r |1 and zero XC kernel ... yields exact exchange and London dispersion forces combine with XC kernels, hybrids with usual XC functionals, split Coulomb interaction . . .

Status of RPA type functionals

Molecular dissociation energies RPA for all e-e separations
error in dissociation energy (eV)

Stacked benzene dimer (Van der Waals complex) short-range: GGA + long-range: RPA
5 4 Interaction energy (kcal/mol) 3 2 1 0 1 2 3 4 3 3.5 4 Separation (A) 4.5 CCSD(T) MP2 vdWDF GGA(revPBE) GGA(PW91)

RPA+
1 0.5 0

RPA

PBE0 hyb
-0.5 -1 -2

PBE GGA
2

O2

F 2 Si 2

2 O O O2 HF C C C H H 2 2

... no consistent improvement over GGA ... not too bad, without X and C error cancellation ... better TD-DFT XC kernels needed!

... includes dispersion forces!

Furche, PRB 64, 195120 (2002), Fuchs and Gonze, PRB 65, 235109 (2002)

Dion et al., PRL 92, 246401 (2004)

Summary
LDA & GGA are de facto controlled approximations to the average XC hole GGA remedies LDA shortcomings w.r.t. total energy dierences but may also overcorrect (e.g. lattice parameters) still can & should check GGA induced corrections for plausibility by . . . simple arguments like homogeneity & coordination . . . results from quantum chemical methods (Quantum Monte Carlo, CI, . . . ) . . . depends on actual GGA functional hybrid functionals mix in exact exchange (B3LYP, PBE0, ... functionals) orbital dependent, implicit density functionals: exact Kohn-Sham exchange, Meta-GGA & OEP method, functionals from the adiabatic-connection uctuation-disspation formula Always tell what XC functional you used, e.g. PBE-GGA (not just GGA) ... helps others to understand your results ... helps to see where XC functionals do well or have a problem