From electrons to photons: Quantum-

inspired modeling in nanophotonics

Steven G. Johnson, MIT Applied Mathematics

 Nano-photonic media (-scale)
strange waveguides

& microcavities
[B. Norris, UMN] [Assefa & Kolodziejski,
MIT]
3d
structures

[Mangan,
Corning]

synthetic materials
optical phenomena
hollow-core fibers
 Photonic Crystals
periodic electromagnetic media
1887 1987
1-D 2-D 3-D

periodic in periodic in periodic in

one direction two directions three directions

can have a band gap: optical “insulators”

 Electronic and Photonic Crystals
atoms in diamond structure dielectric spheres, diamond lattice
Medium
Periodic

photon frequency
Band Diagram
Bloch waves:

electron energy

wavevector wavevector
interacting: hard problem non-interacting: easy problem
 Electronic & Photonic Modelling
Electronic Photonic
• strongly interacting • non-interacting (or weakly),
—tricky approximations —simple approximations
(finite resolution)
—any desired accuracy

• lengthscale dependent • scale-invariant

(from Planck’s h) —e.g. size 10   10

Option 1: Numerical “experiments” — discretize time & space … go

Option 2: Map possible states & interactions
using symmetries and conservation laws: band diagram
 Fun with Math
r r 1∂ r ω r
∇ ×E =− H =i H First task:
c ∂t c get rid of this mess
r r 1∂ r r0 ω r
∇ ×H =ε E +J =i εE
c ∂t c
dielectric function (x) = n2(x)

1 r ⎛ω ⎞ 2 r
+ constraint
r
∇ × ∇ ×H =⎜ ⎟ H ∇ ⋅ H =0
ε ⎝ c⎠
eigen-operator eigen-value eigen-state
Electronic & Photonic Eigenproblems
Electronic Photonic
⎛ h2 2 ⎞ 1 r ⎛ ω ⎞2 r

⎜− ∇ + V ⎟ψ = Eψ ∇ × ∇ ×H =⎜ ⎟ H
⎝ 2m ⎠ ε ⎝ c⎠
nonlinear eigenproblem simple linear eigenproblem
(V depends on e density ||2) (for linear materials)

—many well-known
computational techniques

Hermitian = real E & , … Periodicity = Bloch’s theorem…

A 2d Model System

dielectric “atom”
=12 (e.g. Si)

square lattice,
period a
a
a
E
TM
H
 Periodic Eigenproblems
if eigen-operator is periodic, then Bloch-Floquet theorem applies:

r r i ( k⋅x−ωt ) r
r r
r
can choose: H(x,t) =e Hkr (x)

planewave
periodic “envelope”

Corollary 1: k is conserved, i.e. no scattering of Bloch wave

r
Corollary 2: Hrk given by finite unit cell,
so  are discrete n(k)
 Solving the Maxwell Eigenproblem
2
Finite cell  discrete eigenvalues n
1 ωn
(∇+ ik ) × ( ∇ + ik ) × H n = 2
Hn
ε c
Want to solve for n(k), constraint: (∇+ ik) ⋅H = 0
& plot vs. “all” k for “all” n,
Graphics
are
QuickTime™
needed
decompressor
toand
seeathis picture.

€
1

where:
0.9

0.8

0.7
H(x,y) ei(kx – t)
0.6

0.5

0.4

€
Photonic Band Gap
0.3

0.2
TM bands
0.1

1 Limit range of k: irreducible Brillouin zone

2 Limit degrees of freedom: expand H in finite basis

3 Efficiently solve eigenproblem: iterative methods

Solving the Maxwell Eigenproblem: 1
1 Limit range of k: irreducible Brillouin zone
—Bloch’s theorem: solutions are periodic in k
M

2π ky
first Brillouin zone  X a
kx
= minimum |k| “primitive cell”

€ reduced by symmetry
irreducible Brillouin zone:

2 Limit degrees of freedom: expand H in finite basis

3 Efficiently solve eigenproblem: iterative methods

Solving the Maxwell Eigenproblem: 2a
1 Limit range of k: irreducible Brillouin zone
2 Limit degrees of freedom: expand H in finite basis (N)

N
H =H(xt ) =∑ hmbm(xt ) ˆ 2
solve: A H =ω H
m=1

2
finite matrix problem: Ah=ω Bh

f g =∫f* ⋅g Aml = bm Aˆ bl Bml = bm bl

3 Efficiently solve eigenproblem: iterative methods

Solving the Maxwell Eigenproblem: 2b
1 Limit range of k: irreducible Brillouin zone
2 Limit degrees of freedom: expand H in finite basis
— must satisfy constraint: (∇ + ik) ⋅H = 0

Planewave (FFT) basis Finite-element basis

constraint, boundary conditions:
H(xt ) =∑ HGe iG⋅xt
€ Nédélec elements
G [ Nédélec, Numerische Math.
35, 315 (1980) ]
constraint: H G ⋅ (G + k) = 0
nonuniform mesh,
uniform “grid,” periodic boundaries,
more arbitrary boundaries,
simple code, O(N log N) [ figure: Peyrilloux et al.,
J. Lightwave Tech.
21, 536 (2003) ]
complex code & mesh, O(N)

€ 3 Efficiently solve eigenproblem: iterative methods

Solving the Maxwell Eigenproblem: 3a
1 Limit range of k: irreducible Brillouin zone
2 Limit degrees of freedom: expand H in finite basis
3 Efficiently solve eigenproblem: iterative methods
2
Ah=ω Bh
Slow way: compute A & B, ask LAPACK for eigenvalues
— requires O(N2) storage, O(N3) time
Faster way:
— start with initial guess eigenvector h0
— iteratively improve
— O(Np) storage, ~ O(Np2) time for p eigenvectors
(p smallest eigenvalues)
Solving the Maxwell Eigenproblem: 3b
1 Limit range of k: irreducible Brillouin zone
2 Limit degrees of freedom: expand H in finite basis
3 Efficiently solve eigenproblem: iterative methods
2
Ah=ω Bh
Many iterative methods:
— Arnoldi, Lanczos, Davidson, Jacobi-Davidson, …,
Rayleigh-quotient minimization
Solving the Maxwell Eigenproblem: 3c
1 Limit range of k: irreducible Brillouin zone
2 Limit degrees of freedom: expand H in finite basis
3 Efficiently solve eigenproblem: iterative methods
2
Ah=ω Bh
Many iterative methods:
— Arnoldi, Lanczos, Davidson, Jacobi-Davidson, …,
Rayleigh-quotient minimization

for Hermitian matrices, smallest eigenvalue 0 minimizes:

2 h' Ah minimize by preconditioned
“variational
theorem” ω =min
0
h h' Bh conjugate-gradient (or…)
 Band Diagram of 2d Model System Graphics
are
QuickTime™
needed
decompressor
toand
seeathis picture.

a (radius 0.2a rods, =12)

1

0.9

frequency  (2πc/a) = a / 
0.8

0.7

0.6

0.5

0.4
Photonic Band Gap
0.3

0.2
TM bands
0.1

irreducible Brillouin zone  X M 

M
r E gap for
k  X TM
n > ~1.75:1
H
 The Iteration Scheme is Important
(minimizing function of 104–108+ variables!)
h' Ah
ω = min
2
0 = f (h)
h h' Bh

Steepest-descent: minimize (h + f) over  … repeat

Conjugate-gradient: minimize (h + d)

€ — d is f + (stuff): conjugate to previous search dirs
Preconditioned steepest descent: minimize (h + d)
— d = (approximate A-1) f ~ Newton’s method
Preconditioned conjugate-gradient: minimize (h + d)
— d is (approximate A-1) [f + (stuff)]
 The Iteration Scheme is Important
Graphics
are
QuickTime™
needed
decompressor
toand
seeathis picture.

(minimizing function of ~40,000 variables)

1000000

100000

10000

1000 JÑ
E E E E E E
100
JÑ
EEE EEEE
EEEEEE no preconditioning
% error

JÑ EEEEEEEE
10 Ñ
J Ñ EEEEEEEEEEEEEEEEEEEEEE
J ÑÑ EEE
EEEEEEE
J ÑÑ EEEEEEE
1 J ÑÑ EEEEE
J J ÑÑÑÑ EEEEEEE
J J J J J ÑÑÑÑÑÑÑÑÑÑÑ EEEEEEE
ÑÑÑÑ EEEEEE
JJ Ñ
ÑÑÑ
Ñ
ÑÑ
ÑÑ
ÑÑ
ÑÑ
Ñ EEEEEEEE
0.1 JJ ÑÑ
ÑÑ
Ñ
ÑÑÑ
ÑÑ
Ñ
ÑÑ EEEEEE
Ñ
ÑÑÑ
ÑÑÑ
Ñ EEEEE
J ÑÑ
Ñ
ÑÑÑÑ EEEEEE
J ÑÑ
ÑÑÑÑ EEEE
0.01 ÑÑÑÑÑ EEE
J ÑÑÑÑ EEEE
ÑÑÑ EEE
J ÑÑÑ
Ñ EEE
0.001 Ñ
ÑÑ EEEE
Ñ
preconditioned J
J
ÑÑÑÑÑ EE
EE
EE
EE
EEE
0.0001
conjugate-gradient J
J
no conjugate-gradient EEE
EEEE
EE
0.00001 J
J
0.000001
1 10 100 1000

# iterations
The Boundary Conditions are Tricky
E|| is continuous

E is discontinuous
(D = E is continuous)

Any single scalar  fails:

(mean D) ≠ (any ) (mean E)
 Use a tensor 
⎛ε ⎞
⎜ ⎟ E
⎜ ε ⎟ ||
⎜ −1 ⎟
⎝ ε−1 ⎠ E
 Graphics
are
QuickTime™
needed
The -averaging is Important
decompressor
toand
seeathis picture.

100

H
H
backwards averaging
10 H H H H H
J
B

B
H H
H
H H H
correct averaging
% error

J
J B B changes order
B no Baveraging
J B
B
1 J J
B
B B of convergence
B
J
J J
B
from ∆x to ∆x2
tensor averaging J J
0.1
J J

0.01
10 100 (similar effects
resolution (pixels/period) in other E&M
numerics & analyses)
 Gap, Schmap? Graphics
are
QuickTime™
needed
decompressor
toand
seeathis picture.

a 0.9

0.8

0.7

frequency 
0.6

0.5

0.4
Photonic Band Gap
0.3

0.2
TM bands
0.1

 X M 

But, what can we do with the gap?

 Intentional “defects” are good

microcavities waveguides (“wires”)

3D Photonic C rystal with Defects

 Intentional “defects” in 2d are
Graphics
QuickTime™
needed
decompressor
toand
seeathis picture.

Graphics
are
(Same computation, with supercell = many primitive cells)
Graphics
are
QuickTime™
QuickTime™
needed needed
seeadecompressor
decompressor
to
and toand
this seeathis picture.
picture.

a
Microcavity Blues
For cavities (point defects)
frequency-domain has its drawbacks:

• Best methods compute lowest- bands,

but Nd supercells have Nd modes
below the cavity mode — expensive

• Best methods are for Hermitian operators,

but losses requires non-Hermitian
 Time-Domain Eigensolvers
(finite-difference time-domain = FDTD)

Simulate Maxwell’s equations on a discrete grid,

+ absorbing boundaries (leakage loss)

• Excite with broad-spectrum dipole ( ) source



signal processing Response is many

complex n [ Mandelshtam,
sharp peaks,
J. Chem. Phys. 107, 6756 (1997) ] one peak per mode

decay rate in time gives loss

 Signal Processing is Tricky
signal processing
complex n
?
Graphics
are
QuickTime™
needed
a common approach: least-squares fit of spectrum
decompressor
toand
seeathis picture. Graphics
are
QuickTime™
needed
decompressor
toand
seeathis picture.

0.8 450
E
0.6 EE 400
EE

0.4 EE
E
350 fit to:
E E
EE
0.2
E
E
EE
E EE E EE
EEEE
E EE
300 E E
A
0 E E EE
E
EEE EE
E E E E E EE EE EEEEE EE
EEEEE
EE
EE
EE
EEE
EEEEEE
EEEE
E
EE
EE
EE
EE
EE
EE
EE
E
EE
EE
EE
EE
EE
EE
E
EE
EE
EE
EE
EE
EE
E
EE
EE
EE
EE
EE
EE
EE
E
EE
EE
EE
EE
EE
EE
E
EE
EE
EE
EE
EE
EE
EE
E
EE
EE
(ω − ω 0 ) 2 + Γ 2
E
EEEE EE
EE 250
E E EE
E E EEE
E EE
-0.2 E
200

FFT
E EE
EE
EE E E
-0.4 150
E
-0.6 100 E E
E
-0.8
E
50
€EE
E
E E
E
EE
-1 E EE E E EE
EE E E E EE E EE EE E EE EE E E
0
0 1 2 3 4 5 6 7 8 9 10 0 0.5 1 1.5 2 2.5 3 3.5 4

Decaying signal (t) Lorentzian peak ()

 Fits and Uncertainty
problem: have to run long enough
Graphics
are
QuickTime™
needed
decompressor
toand
seeathis picture. Graphics
are
totoand
QuickTime™
needed
completely
decompressor
seeathis picture.
decay
1 40000
E
EE
E
0.8 EE
EE E 35000
EE

0.6 EE
EE
EE EEE
EE
EEE
EE EE
E
actual
EE EE
E 30000
EE
EE
EE
EE EE
E EE
0.4 E E EE E
EEE
E E EE EE EE
E E E E EE
E E EE EE 25000
0.2 E E E E E E E E E E E E E E
0 E E E E E E E E E E E E E E E E E E E E 20000

E E E E E E E
-0.2 E E E E E E E
E E EE E
E E E E E E 15000
E EE EE EE E
EE EE E
-0.4 EE
EE EE EEE
EE
E
EE
E E EE EEE
-0.6 E
EE
EE
EE
EE
E
EE
EE
E
EEE
10000
signal
E
EE
EE
-0.8 E
EE
E
EE
E
5000
portion
E E E
-1 E 0 E E E E E E E E
0 1 2 3 4 5 6 7 8 9 10 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5

Portion of decaying signal (t) Unresolved Lorentzian peak ()

There is a better way, which gets complex  to > 10 digits

 Unreliability of Fitting Process
Resolving two overlapping peaks is
near-impossible 6-parameter nonlinear fit
(too many local minima to converge reliably)

sum of two peaks

There is a better
way, which gets
complex 
for both peaks
 = 1+0.033i to > 10 digits
 = 1.03+0.025i

Sum of two Lorentzian peaks ()

 Quantum-inspired signal processing (NMR spectroscopy):
Filter-Diagonalization Method (FDM)
[ Mandelshtam, J. Chem. Phys. 107, 6756 (1997) ]

Given time series yn, write: y n = y(nΔt) = ∑ ak e −iω k nΔt

…find complex amplitudes ak & frequencies k

by a simple linear-algebra problem!
Idea: pretend y(t)
€ is autocorrelation of a quantum system:
ˆ ∂ ˆ −iHˆ Δt / h
H ψ = ih ψ time-∆t evolution-operator: U = e
∂t
say: y n = ψ (0) ψ (nΔt) = ψ (0) Uˆ n ψ (0)
Filter-Diagonalization Method
[ Mandelshtam, J. Chem. Phys. 107, 6756 (1997) ]
(FDM)

y n = ψ (0) ψ (nΔt) = ψ (0) Uˆ n ψ (0) ˆ

U =e −iHˆ Δt / h

We want to diagonalize U: eigenvalues of U are ei∆t

€
…expand U in basis of |(n∆t)>:

U m,n = ψ (mΔt) Uˆ ψ (nΔt) = ψ (0) Uˆ mUˆ Uˆ n ψ (0) = y m +n +1

Umn given by yn’s — just diagonalize known matrix!

 Filter-Diagonalization Summary
[ Mandelshtam, J. Chem. Phys. 107, 6756 (1997) ]

Umn given by yn’s — just diagonalize known matrix!

A few omitted steps:
—Generalized eigenvalue problem (basis not orthogonal)
—Filter yn’s (Fourier transform):
small bandwidth = smaller matrix (less singular)

• resolves many peaks at once

• # peaks not known a priori
• resolve overlapping peaks
• resolution >> Fourier uncertainty
 Do try this at home

Bloch-mode eigensolver:
http://ab-initio.mit.edu/mpb/

Filter-diagonalization:
http://ab-initio.mit.edu/harminv/

Photonic-crystal tutorials (+ THIS TALK):

http://ab-initio.mit.edu/
/photons/tutorial/