Introduction to the calculation of phonons

and of vibrational spectra

P. Giannozzi
Università di Udine and Democritos National Simulation Center, Italy

September 23, 2009

– Typeset by FoilTEX –
 Vibrational Spectroscopy: experiments

• Observation of vibrational modes:

– phonons in crystals,
– normal modes in molecules,
is a powerful tool in materials characterization

• Vibrational spectroscopy is a sensitive probe of the atomic structure and of the

chemical bonding and thus of the electronic structure

• Most frequently used experimental techniques:

– Neutron scattering (technically difficult, the entire dispersion is observable)
– Infrared (IR) spectroscopy (simple, only some modes are observable)
– Raman spectroscopy (simple, a different subset of modes is observable)

Theoretical calculation of vibrational mode frequencies and intensities from first

principles is very helpful in analyzing vibrational spectra
 Basic theory: Born-Oppenheimer approximation

The behavior of a system of interacting electrons r and nuclei R is determined by

the solutions of the time-dependent Schrödinger equation:
!
2
X h̄ ∂ 2 X h̄ ∂ 2 2
∂ Φ̂(r, R; t)
ih̄ = − 2 − 2 + V (r, R) Φ̂(r, R; t)
∂t 2MI ∂RI i
2m ∂ri
I

where V (r, R) is the potential describing the coulombian interactions:

X ZiZJ e2 X ZI e2 X e2
V (r, R) = − +
|RI − RJ | |ri − RI | i>j |ri − rj |
I>J i,I

≡ Vnn(R) + Vne(r, R) + Vee(r)

Born-Oppenheimer (or adiabatic) approximation (valid for MI >> m):

Φ̂(r, R; t) ' Φ(R)Ψ(r|R)e−iÊt/h̄

NB: r ≡ (r1, .., rN ), R ≡ (R1, .., Rn)

 Basic theory: Potential Energy Surface

The Born-Oppenheimer approximation allows to split the problem into an electronic

problem depending upon nuclear positions:
!
X h̄ ∂2 2
− 2 + V (r, R) Ψ(r|R) = E(R)Ψ(r|R)
i
2m ∂ri

and a nuclear problem under an effective interatomic potential determined by the

electrons: !
2
X h̄ ∂ 2
− + E(R) Φ(R) = ÊΦ(R)
2MI ∂R2i
I
E(R) determines the Potential Energy Surface and the equilibrium geometry. At
equilibrium, forces FI on nuclei vanish:

∂E(R)
FI = − =0
∂RI

NB: r ≡ (r1, .., rN ), R ≡ (R1, .., Rn)

 Normal vibrational modes in molecules and crystals
Harmonic approximation: the interatomic potential energy is expanded to 2nd
order. The resulting Hamiltonian transforms into a sum of independent oscillators.
Normal mode frequencies, ω, and displacement patterns, UIα for cartesian
component α of atom I, at atomic position RI , are determined by the secular
equation: X  αβ 
CIJ − MI ω δIJ δαβ UJβ = 0,
2

J,β
αβ
where CIJ is the matrix of inter-atomic force constants, i.e. second derivatives of
the energy with respect to atomic positions:

αβ ∂ 2E({R}) ∂FαI ({R})

CIJ ≡ =−
∂RIα∂RJβ ∂RJβ

In crystals, normal modes are classified by a wave-vector q. Phonon frequencies,

ω(q), and displacement patterns, Usα(q), are determined by the secular equations:
X  αβ 
Cest (q) − Msω 2(q)δstδαβ Utβ (q) = 0
t,β
 Practical calculation of vibrational modes
Common methods – based on Density-functional Theory – to calculate vibrational
modes (via Interatomic Force Constants):

• Frozen-phonon technique: finite differentiation of forces

• Density-functional Perturbation Theory (DFPT): direct calculation of second-

order derivatives of the energy

Alternative method (not using Interatomic Force Constants):

• Molecular Dynamics: spectra can be extracted from Molecular Dynamics runs,

via the Fourier Transform of the velocity-velocity autocorrelation function
Z ∞ X hvI (t) · vI (0)i
1 −iωt
f (ω) = √ e dt, (1)
2π −∞ hvI (0) · vI (0)i
I

Function f (ω) has peaks at the vibrational frequencies.

Let us focus on the case of phonons in crystals.

 Calculation of phonon spectra
Introduce monochromatic perturbation u to atomic positions RI = Rl + τ s as

RI [us(q)] = Rl + τ s + us(q)eiq·Rl .

(Rl =lattice vector, τ s =equilibrium position of the s-th atom in the unit cell).
Fourier transform of force constants at q are second derivatives of the energy with
respect to such monochromatic perturbations:

αβ
X
−iq·R αβ 1 ∂ 2E
C
est (q) ≡ e Cst (R) = β
Nc ∂u∗α
s (q)∂u t (q)
R

Quantum ESPRESSO calculates them from the knowledge of the linear response
∂n(r)/∂uα
s (q) and diagonalized to get phonon modes at q. Note that:

• the linear response has the same wave vector q of the perturbation: this
algorithm will work for any q without any supercell involved

• in the spirit of adiabatic approximation, one can use static response.

 Frozen phonon

Frozen phonons is an older and alternative way to calculate phonons. The

monochromatic perturbation is frozen in with a finite amplitude in the system,

Fourier transform of force constants at q are calculated from finite differences of

forces induced on all the atoms of the supercell by the monochromatic perturbation.

Advantages:

• straightforward to implement

Disadvantages:

• limited to small supercells, i.e. q = G/n, where G is a reciprocal lattice vector

of the original cell, n = 2, 3, 4, ..., but in any case a small number.

Note that this is not the algorithm used by Quantum ESPRESSO!

What if we want the entire dispersions for all q-vectors in the Brillouin Zone?
 Calculation of interatomic force constants

αβ
Inter-atomic force constants in real-space, Cst (R), are obtained by

αβ
• calculating C
est (q) on a discrete (n1, n2, n3) grid of q-vectors:

i−1 j−1 k−1

qijk = G1 + G2 + G3 , i = 1, .., n1, and the like for j, k;
n1 n2 n3

• Fourier-transforming to the corresponding real-space grid:

C(qijk ) ⇐⇒ C(Rlmn), Rlmn = lR1 + mR2 + nR3

l = −n1/2, ..., n1/2 and the like for m, n.

The denser the grid of q-vectors, the larger the vectors Rlmn for which the
inter-atomic force constants are calculated. Inter-atomic force constants are short-
ranged and require a moderate number of calculations at different q (for polar
systems, see later).
 Density-Functional Perturbation Theory
Let us assume that the external potential depends on some parameter λ

∂V (r) 1 2 ∂ 2V (r)
Vλ(r) ' V (r) + λ + λ 2
+ ...
∂λ 2 ∂λ
(all derivatives calculated at λ = 0) and expand the charge density

∂n(r) 1 2 ∂ 2n(r)
nλ(r) ' n(r) + λ + λ + ...
∂λ 2 ∂λ2
and the energy functional into powers of λ:

∂E 1 2 ∂ 2E
Eλ ' E + λ + λ + ...
∂λ 2 ∂λ2
The first-order derivative ∂E/∂λ does not depend on any derivative of n(r)
(Hellmann-Feynman theorem):
Z
∂E ∂V (r)
= n(r) dr
∂λ ∂λ
 Density-Functional Perturbation Theory II

The second-order derivative ∂ 2E/∂λ2 depends on the first-order derivative of the

charge density, ∂n(r)/∂λ:

∂ 2E ∂ 2V (r)
Z Z
∂V (r) ∂n(r)
= dr + n(r) dr
∂λ2 ∂λ ∂λ ∂λ2

The result can be generalized to mixed derivatives:

∂ 2E ∂ 2V (r)
Z Z
∂V (r) ∂n(r)
= dr + n(r) dr
∂λ∂µ ∂λ ∂µ ∂λ∂µ

(the order of derivatives can be exchanged)

In general, the (2n + 1)−th derivative of energy depends only on derivatives up to

order n of the charge density ((2n + 1) theorem) due to its variational character.

∂n/∂λ can be calculated either by a self-consistent procedure, or by direct

minimization of the 2nd-order energy, written as a functional of ∂n/∂λ.
 Calculation of the Linear Response
The minimization of the second-order functional, or alternatively the linearization
of Kohn-Sham equations, yields a set of equations:

∂ψi(r) X ∂ψj ∂V
(HKS − i) + Kij = −Pc |ψii
∂λ j
∂λ ∂λ

where Pc is the projector over unoccupied states, Kij is a nonlocal operator:

2
  Z  
∂ψj e δvxc(r) ∗ 0 ∂ψj 0 0
Kij (r) = 4 ψi(r) + ψ j (r ) (r )dr
∂λ |r − r0| δn(r0) ∂λ

and the linear charge response:

∂n(r) X
∗ ∂ψi(r)
= 2Re ψi (r)
∂λ i
∂λ

These equations can be solved as a linear system for all ∂ψi(r)/∂λ, or

else as many linear systems for each ∂ψi(r)/∂λ followed by self-consistency
 Phonons and macroscopic electric fields
Polar materials in the q=0 limit: a macroscopic electric field appear as a
consequence of long-rangeness of Coulomb interactions. Incompatible with Periodic
Boundary Conditions! Macroscopic electric fields contribute a non-analytic term
to the dynamical matrix:
? ?
na e αβ 4π (q · Z s)α (q · Z t)β
Cst =
Ω q · ∞ · q

Effective charges Z? are related to polarization P induced by a lattice distortion:

∂Pα
Z ?αβ
s =Ω
∂uβs (q = 0)

Dielectric tensor αβ

∞ are related to polarization induced by an electric field E:

∂Pα 
αβ
∞ = δαβ + 4π .
∂Eβ  us (q=0)=0

All of the above can be calculated from (mixed) second derivatives of the energy.
 Practical calculations for electric fields

We need the response to a macroscopic electric field E: ∂V /∂Eα = −erα.

This is ill-defined in a crystal, because r is not a lattice-periodic operator!
It can however be recast into a well-defined expression using the following trick:

hψc|[HKS , r]|ψv i
hψc|r|ψv i = for c 6= v.
c − v

We can rewrite |ψ̄vαi = Pcrα|ψv i as the solution of a linear system:

(HKS − v )|ψ̄vαi = Pc[HKS , rα]|ψv i,

where the commutator can be calculated from the following expression:

h̄2 ∂ h i
[HKS , r] = − + V̂N L, r .
m ∂r

(VN L is the nonlocal term of the potential if present).

Alternative technique: differentiation wrt Bloch vector (numerically performed)
 Calculation of Infrared Cross Sections

Infrared spectroscopy (absorption/reflectivity, energy loss spectroscopy, FTIR):

λ >> a, the electromagnetic field can be considered macroscopic. Selection rules:

• Only phonons close to the Γ-point (q = 0) are involved, i.e. IR-active

• Only phonons that induce a nonzero dipole are IR-active

The IR intensity is proportional to the square of the dipole induced by the phonon
that is excited by the IR radiation:
 2
 
X X αβ β 
?
IIR(ν) = 
 Z I UI (ν) .

α  Iβ 

This is a trivial byproduct of a phonon calculation.

Alternative technique: effective charges are expressed via the Berry’s phase.
 Calculation of Raman Cross Sections

Non-resonant, resonant, surface-enhanced Raman scattering: an inelastic process

in which light is scattered and a phonon or normal mode is created or destroyed

Again, only phonons near Γ (k = 0) contribute.

Stokes process: ωS = ωL − ωk
Anti-Stokes process: ωS = ωL + ωk

Resonant Raman: ωL,S ∼ ωel = typical electronic excitation frequencies

Non-resonant Raman: ωL,S << ωel. This is the simpler case.
 Calculation of Raman Cross Sections (2)

Non-resonant Raman intensities in the Placzek approximation:

4  ∂χαβ 2
 
(ωL − ων )
IStokes(ν) ∝ rαβ (ν), rαβ (ν) =  (ν) 
ων ∂u

χ = electric polarizability, u(ν) = normal mode coordinate along mode ν. The

Raman tensor rαβ (ν) is a third-order derivative of the energy that can be calculated
using:

• DFPT for χ + Frozen-phonon (finite differences)

• DFPT using the (2n+1) theorem: the (2n+1)−th derivative of energy depends
only on derivatives up to order n of the charge density (not implemented).

• DFPT + Second-order response to electric field (imlpemented)

• Finite electric fields + Frozen phonon

 Practical phonon calculation in Quantum ESPRESSO

First step: scf calculation at equilibrium positions (performed by pw.x)

• Single phonon calculation at finite wave-vector q

– Generate ψk,v and ψk+q,v in the Irreducible Brillouin Zone relative to the
small group of q; calculate C(q), diagonalize, produce ω(q) and U (q) (code
ph.x)

• Single phonon calculation at Γ wave-vector (q=0)

– Calculate C(q = 0), diagonalize, produce ω(q = 0) and U (q = 0) (ph.x)
For polar materials: calculate non-analytical terms that are missing from
C(q = 0) (LO-TO splitting are absent from ω(q = 0)): specify option
epsil=.true. to ph.x (will calculate and store in output file Z ∗ and ∞).
For non resonant Raman: specify option lraman=.true.
– Impose Acoustic Sum Rule (ASR), add the nonanalytic LO-TO splitting,
calculate cross sections (if data available) (code dynmat.x)
 Practical phonon dispersions calculation

First step as before: scf calculation at equilibrium positions (performed by pw.x)

• Perform many single-phonon calculations on a uniform grid of wave-vectors qi,

including q = 0 (if system is polar, calculate in the latter case Z ∗ and ∞); save
all C(q1) (and Z ∗, ∞) (code ph.x with option ldisp=.true.)

• Perform inverse FFT of the C(qi), obtain interatomic force constants in real
space C(R). For polar materials: a term having the same behaviour for q → 0
as the non-analytic term is subtracted from C(qi) before the Fourier Transform
and re-added to C(R), so that no problem related to non-analytic behaviour
and related long-rangeness arises in the Fourier Transform (code q2r.x)

• Calculate phonons at any wave-vector using code matdyn.x

 Fast algorithm for specific cases

If you sample the Brillouin Zone with only the Γ point (e.g. molecules, large unit
cells) and you need phonon modes only at Γ, you can use a simplified and faster
algorithm.

• scf calculation at equilibrium positions with Γ−point tricks: performed by pw.x

with card
K POINTS gamma

• use specialized code phcg.x to find C(q = 0); specify option epsil=.true.
to calculate Z ∗ and ∞.

• Impose Acoustic Sum Rule (ASR), add the nonanalytic LO-TO splitting,
calculate IR cross sections with code dynmat.x

Restrictions: no Raman, no Ultrasoft Pseudopotentials