Efficient Solution of the Schrödinger-Poisson

Equations in Semiconductor Device Simulations

Alex Trellakis, Till Andlauer, and Peter Vogl

Walter-Schottky Institute, TU München, D-85748 Garching, Germany

Abstract. This paper reviews the numerical issues arising in the sim-
ulation of electronic states in highly confined semiconductor structures
like quantum dots. For these systems, the main challenge lies in the
efficient and accurate solution of the self-consistent one-band and multi-
band Schrödinger-Poisson equations. After a brief introduction of the
physical background, we first demonstrate that unphysical solutions of
the Schrödinger equation due to the presence of material boundaries can
be avoided by combining a suitable ordering of the differential operators
with a robust discretization method like box discretization. Next, we dis-
cuss algorithms for the efficient solution of the resulting sparse matrix
problems even on small computers. Finally, we introduce a predictor-
corrector-type approach for the stabilizing the outer iteration loop that
is needed to obtain a self-consistent solution of both Schrödinger’s and
Poisson’s equation.

1 Physical Background

1.1 Introduction

Due to the rapid progress in semiconductor technology, the dimensions of elec-

tronic devices are approaching the nanometer scale. At this scale, electronic
properties will be increasingly dominated by quantum effects, until at some point
these effects become essential for device operation. For instance, the electronic
states in highly confined semiconductor structures like quantum dots are entirely
controlled by quantum mechanical size quantization, which can be exploited for
entirely new device designs like quantum well or quantum dot lasers. At the same
time, the physical equations that describe these quantum systems are much too
complex to be solved analytically and efficient numerical solution methods need
to be used instead. For this reason, we review in this paper the numerical is-
sues arising in the simulation of electronic states in confined quantum systems,
with focus on the efficient solution of the self-consistent Schrödinger-Poisson
equations.

1.2 The k·p (Schrödinger) Equations

Since these quantum structures are about one to two orders of magnitude larger
than a single lattice constant (which is about 0.5 nm), atomistic models like

I. Lirkov, S. Margenov, and J. Waśniewski (Eds.): LSSC 2005, LNCS 3743, pp. 602–609, 2006.

c Springer-Verlag Berlin Heidelberg 2006
 Efficient Solution of the Schrödinger-Poisson Equations 603

empirical tight binding theory or empirical pseudopotential methods are numer-

ically not feasible and a continuum model with carefully tuned material parame-
ters needs to be used. Such a continuum model is provided by the k·p-equations
in the envelope function approximation [10, 5, 2].
In this model, we describe the electrons in the conduction band by an one-
band k·p-equation which is called the effective-mass approximation. Here, the
electron wave functions ψne (x) obey the Schrödinger equation

Ene ψne (x) = Ĥeff ψne (x) (1)

with the effective mass Hamiltonian


1 1

Ĥeff = p̂ · p̂ + EC (x̂) + V̂xc (x̂) p̂ = ∇ (2)
2 m∗ i

where Ene is the energy of wave function ψne (x), EC (x) the conduction band
edge, and Vxc (x) the exchange-correlation potential in the local density approx-
imation. Finally, m∗ is a 3 × 3-tensor describing the effective electron mass in
the semiconductor.
For holes in the valence bands, the effective-mass approximation is usually not
accurate enough anymore, and a multi-band approach treating the three main
valence bands simultaneously is needed. Since each of the three valence bands
has two spin components, we need here the 6-band k·p-equation

Enh ψnh (x) = Ĥ6×6 ψnh (x) (3)

where ψnh (x) is a 6-component hole wave function, and Ĥ6×6 a Hermitian 6 × 6-
matrix operator


Ĥ3×3 0 6×6
Ĥ6×6 = + Ĥso (4)
0 Ĥ3×3
6×6
Here, Ĥso is a momentum independent complex matrix that describes the spin-
orbit interaction, and Ĥ3×3 a 3×3-matrix that for a Zincblende crystal structure
oriented along the coordinate axes has the form [9]
    1 2 N
Ĥ3×3 = EV (x̂) + V̂xc (x̂) δij + p̂ δij + 2 p̂i p̂j
ij 2m0

L−N −M 2 M
+ p̂i δij + 2 p̂2 δij (5)

2

where EV (x) is the valence band edge, m0 the vacuum mass of the electron, and
L, M, and N the Dresselhaus parameters for the semiconductor [5].
Since the focus of our simulations are closed systems, we use homogeneous
Dirichlet conditions (ψ ≡ 0) for all resulting Schrödinger equations. But note
that the energies and wave functions of bound states in a confining potential
are only weakly affected by their boundary conditions, as long as the simulation
region in which Schrödinger’s equation is solved is large enough.
604 A. Trellakis, T. Andlauer, and P. Vogl

1.3 Carrier Densities and the Poisson Equation

In thermal equilibrium, the electron and hole densities n (x) and p (x) are cal-
culated from the quantum states {En , ψn } as [15]

e 
2 En − EF
n (x) = 2 |ψne (x)| F (6)
n
kB T



ψnh (x)2 F EF − En
h
p (x) = 2 (7)
n
kB T

where kB is the Boltzmann constant, T the temperature of the system, and F (E)
a suitable distribution function. This F (E) is in three dimensions equal to the
Fermi function
1
F (E) = (8)
1 + exp (E)
while in one and two dimensions F (E) is proportional to the complete Fermi-
Dirac integrals F0 (−E) and F−1/2 (−E). Furthermore, in addition to these car-
rier densities there may also be space charges

ND+ (x) = ND+ [EC (x̂)] NA− (x) = NA− [EV (x̂)] (9)

present due to doping with electron donors (ND+ ) and acceptors (NA− ).
All these charge densities add up to a total charge

ρ (x) = e p (x) − n (x) + ND+ (x) − NA− (x) (10)

which then results through Poisson’s equation

∇ · [ (x) ∇φ (x)] = −ρ (x) (11)

into an electric potential φ (x) that warps the band edges EC and EV as

EC (x) −→ EC [φ] (x) = EC (x) − eφ (x) (12)

EV (x) −→ EV [φ] (x) = EV (x) − eφ (x) (13)

As consequence, the charge density ρ becomes potential dependent and the Pois-
son equation (11) nonlinear in φ
 
∇ · (∇φ) = −e p − n + ND+ [φ] − NA− [φ] (14)

Note that the carrier densities n and p depend here indirectly on φ, since the
band edges EC [φ] and EV [φ] enter the Hamiltonian, which of course makes
also all wave functions and energies potential dependent. As a consequence,
the nonlinear Poisson equation (14) and the Schrödinger equations become a
coupled system of differential equations for which we need to find a self-consistent
solution.
 Efficient Solution of the Schrödinger-Poisson Equations 605

1.4 Operator Ordering and Symmetrization

An important issue we have ignored so far is that the material parameters in

(2) and (5) are position dependent in heterostructures. This is a major prob-
lem since both Hamiltonians are strictly speaking only valid for a homogeneous
semiconductor with slowly varying band edges EC (x) and EV (x). Some numer-
ical experimentation now shows that even in the presence of material induced
discontinuities in the band edges these Hamiltonians still deliver physically rea-
sonable results. This empirical observation can also be partially justified using
Burt’s exact envelope function theory [3, 4].
Unfortunately, variations in m∗ and L, M, N are much more problematic, since
here we need also to define the operator ordering between these material para-
meters and the momentum operator p̂. While it is clear that any symmetrization
procedure used must yield a Hermitian Hamiltonian
†
H (x̂, p̂) = H (x̂, p̂) (15)
this requirement is not sufficient for uniquely defining H (x̂, p̂). Furthermore, a
poorly chosen H (x̂, p̂) will result in wave functions that become unphysical at
material interfaces and lead to highly incorrect solutions [7].
For the effective mass approximation (2), we can use in analogy to Poisson’s
equation (11) the Hamiltonian [11]

1 1
Heff (x̂, p̂) = p̂ · p̂ + EC (x̂) + V̂xc (x̂) (16)
2 m∗ (x̂)
where compared to (2) we have just replaced the effective mass m∗ by the opera-
tor m∗ (x̂). This Hamiltonian is flux conserving and leads to wave functions that
are continuous at material interfaces and smooth everywhere else. Since most
alternative Hamiltonians do not share these properties [11], Hamiltonian (16) is
widely used in quantum simulations.
For the 6-band k·p-Hamiltonian (4-5) finding a suitable symmetrization is
more difficult since Ĥ6×6 as 6 × 6-matrix requires only
 †  
Ĥ6×6 = Ĥ6×6 (17)
ij ji

in order to become Hermitian. Here, one can show that starting again from Burt’s
exact envelope function theory a suitable 6-band k·p-Hamiltonian Ĥ6×6 (x̂, p̂)
can be constructed using the rather complicated operators [7, 8]
  1 2 N (x̂)
[H3×3 (x̂, p̂)]ij = EV (x̂) + V̂xc (x̂) δij + p̂ δij + p̂i 2 p̂j
2m0

 
L (x̂) − N (x̂) − M (x̂) M (x̂)
+ p̂i p̂ i δ ij + p̂ · p̂ δij

2
2
M (x̂) M (x̂)
− p̂i p̂j + p̂j p̂i (18)

2
2
where the two last terms vanish in a bulk semiconductor. The resulting Hamil-
tonian H6×6 (x̂, p̂) is again flux conserving and yields wave functions that are
continuous at material interfaces and smooth everywhere else.
606 A. Trellakis, T. Andlauer, and P. Vogl

2 Numerical Issues
2.1 Discretization
Both the Poisson equation and the Schrödinger equations are discretized using
box integration on a nonuniform tensor product grid. Such a grid can be seen
as the tensor product of D nonuniform one-dimensional grids {xi } along each
coordinate axis, where D is the dimension of the simulation domain. Such tensor
grids have the advantage that no complicate meshing algorithm is required, and
the discretization can be easily and efficiently implemented. In our implemen-
tation, the user provides the desired density of grid lines at selected locations
along the coordinate axes. The resulting tensor product grid is then calculated
by interpolating these line densities.
For the box integration finite difference method [14–page 184], we cover the
simulation domain with N non-overlapping rectangular boxes, where each box
is centered on a grid node. We then integrate our differential equations over each
box volume Ω to obtain an integral equation which can then be discretized. For
instance, the Poisson equation (14) becomes after integration over Ω
 
n(x) · [(x)∇φ(x)] = − ρ(x) (19)
∂Ω Ω

which after discretization then yields the familiar three, five, or seven-point sten-
cils for φ.
Compared to standard finite differences, box integration has the advantage
that discontinuities in a material parameter A are naturally taken into account
for second order differential operators ∂ˆi A (x̂) ∂ˆj , since this method is in first
order flux conservative due to the Gauss theorem. This makes box integration
a natural choice for discretizing the Poisson equation (14) and the Schrödinger
equations resulting from (16) and (18).
After the discretization, all operators ∂ˆi A (x̂) ∂ˆj in these equations become
banded sparse matrices. Such matrices can be easily stored and manipulated
using banded storage schemes that store only the non-zero diagonals. In specific,
the Schrödinger equation belonging to the Hamiltonians (16) or (18) becomes
after discretization a generalized symmetric matrix eigenvalue problem of the
form
Hψn = En Dψn Dij = δij Vol (Ωi ) (20)
with D being a diagonal matrix containing the volumes of the box regions Ωi
on the main diagonal. Similarly, the Poisson equation (14) as a boundary value
problem yields a linear system, if we ignore the dependence of the charge density
ρ (φ) on the potential φ for now.

2.2 Solving the Nonlinear Poisson Equation

At this point, solving each of the two differential equations offers its own set of
challenges. For solving the nonlinear Poisson equation (14)
f (φ) ≡ ∇ · (∇φ) + ρ(φ) = 0 (21)
 Efficient Solution of the Schrödinger-Poisson Equations 607

we can immediately use Newton-Raphson iteration with inexact line search as

−1
φ̃k (λ) ≡ φk − λ [(Dφ f ) (φk )] · f (φk ) (22)
  
  1 1 1
φk+1 ≡ φ̃k (λ) with f φ̃k (λ)  ≡ min λ = 1, , , , . . . (23)
2 4 8
in order to find a solution of the nonlinear problem (experience shows that local
minima in f (φ) or multiple solutions do not occur for nanostructures).
We also note that, except for the extra diagonal term Dφ ρ, the Jacobian
matrix in (22)
 · (∇)
(Dφ f ) (φ) = ∇  + (Dφ ρ) (φ) (24)
 · (∇).
is identical to the Poisson operator ∇  Since ∇  · (∇)
 is positive definite
and Dφ ρ is positive for all physical ρ(φ), the Jacobian Dφ ρ must be positive
definite as well. Therefore, we can employ the preconditioned conjugate gradient
method [12–chapter 6] for solving the linear system in (22). As preconditioner we
use here the very fast fill-in free Dupont-Kendall-Rachford method [6] in order
to achieve rapid convergence of the conjugate gradient iteration.

2.3 Solving the Schrödinger Equations

Solving the matrix eigenvalue systems resulting from the discretization of the
Schrödinger equations is computationally much more demanding. Especially the
6-band k·p-Hamiltonian (18) results here into extremely large sparse matrices
that prohibit the use of standard diagonalization routines. Luckily enough, all
distribution functions F (E) that are used in calculating the electron and hole
densities (6) and (7) fall off exponentially as

F (E) ∝ exp (−E) , E  0. (25)

For this reason, only relatively few quantum states at the lower end of the energy
spectrum are needed to calculate the densities, which allows us to use highly
efficient iterative eigenvalue methods like the Lanczos or Arnoldi iteration [12].
One software package that implements such iteration methods for both real
and complex matrices is ARPACK [1]. In the authors’ experience, ARPACK
appears quite reliable and moderately fast for the matrices that occur in the
simulation of quantum structures. However, ARPACK is still not fast enough for
the huge system matrices that arise in three-dimensional quantum simulations
and an additional preconditioner is needed.
Such a preconditioner can be easily obtained by utilizing the Chebyshev poly-
nomials Tn (x) for a spectral transformation. Here we use that if the Hermitian
matrix Ĥ has the spectrum {En }, the matrix T (Ĥ) must have the spectrum
{T (En )}. Since the polynomials Tn (x) increase very rapidly for |x| > 1, we can
use these Tn (x) to make the desired eigenvalues at the lower end of the spec-
trum very large, which accelerates the convergence of the Arnoldi iteration in
ARPACK tremendously.
608 A. Trellakis, T. Andlauer, and P. Vogl

As a quick test for this preconditioned ARPACK code, we solve the eigenvalue
problem for test matrices obtained by discretizing an one-band problem (16). For
these matrices, we get for a 3 GHz PC and 15-th order Chebyshev polynomials
as solution times

grid size # states CPU time

50 × 50 × 50 30 17 s
50 × 50 × 50 200 357 s
100 × 100 × 100 30 269 s

This test shows that memory exhaustion now surpasses excessive CPU time
as the main concern for quantum simulations.

2.4 Solution of the Coupled System

As the last step, we need to tackle the self-consistent solution of the coupled
nonlinear Poisson and Schrödinger equations. Here, a simple iteration does not
converge due to the strong coupling between the equations. Similarly, under-
relaxation stabilizes the outer iteration only very poorly, and we still observe
strong charge oscillations from one iteration step to the next that interfere with
convergence [13].
This situation can be much improved by partially decoupling both partial
differential equations using a predictor-corrector-type approach. In order to do
this, we replace the exact carrier densities n (x) and p (x) from (6) and (7) by
the φ-dependent predictors [13]

e 
En − EF + e [φ (x) − φprev (x)]
ñ [φ] (x) = 2 |ψne (x)|2 F (26)
n
kB T



ψnh (x)2 F EF − En − e [φ (x) − φprev (x)]
h
p̃ [φ] (x) = 2 (27)
n
kB T

where φprev is the electrostatic potential from the previous outer iteration step.
These predictors for the quantum densities n and p are then used in the nonlinear
Poisson equation
 
∇ · (∇φ) = −e p̃ [φ] − ñ [φ] + ND+ [φ] − NA− [φ] (28)

in order to determine the new potential φ (x). Using this φ (x) we update the
band edges EC (x) and EV (x) in the Hamiltonians (16) and (18), and solve the
respective Schrödinger equations for a new set of energies and wave functions.
Obviously, once the iteration has converged we have φ = φprev and therefore
with n = ñ [φ] and p = p̃ [φ] also the correct densities. The numerical experiment
then shows [13] that this approach leads to rapid convergence with the residuals
in the quantum densities n and p decreasing by about one order of magnitude.
No further steps are necessary to ensure convergence [13].
 Efficient Solution of the Schrödinger-Poisson Equations 609

References
1. ARPACK libraries and related publications http://www.caam.rice.edu/
software/ARPACK/.
2. Bastard, G.: Superlattice band structure in the envelope-function approximation.
Phys. Rev. B 24 (1981) 5693–5697
3. Burt, M.G.: The justification for applying the effective-mass approximation to
semiconductors. J. Phys.: Condens. Matter 4 (1992) 6651–6690
4. Burt, M.G.: Fundamentals of envelope function theory for electronic states and
photonic modes in nanostructures. J. Phys.: Condens. Matter 11 (1999) R53–R83
5. Dresselhaus, G., Kip, A.F., Kittel, C.: Cyclotron Resonance of Electrons and Holes
in Silicon and Germanium Crystals. Phys. Rev. 98 (1955) 368–384
6. Dupont, T., Kendall, R.P., Rachford, H.H.: An approximate factorization proce-
dure for solving self-adjoint elliptic difference equations. SIAM J. Numerical Analy-
sis 5 (1968) 559–573
7. Foreman, B.A.: Effective-mass Hamiltonian and boundary conditions for the va-
lence bands of semiconductor microstructures. Phys. Rev. B 48 (1993) 4964–4967
8. Foreman, B.A.: Elimination of spurious solutions from eight-band k·p theory. Phys.
Rev. B. 56 (1997) R12748–12751
9. Kane, E.O.: The k·p method. In: Semiconductors and Semimetals: III–V com-
pounds, vol. 1, ed by R.K. Willardson, A.C. Beer (Academic Press, New York
1966) 75–100
10. Luttinger, J.M., Kohn, W.: Motion of Electrons and Holes in Perturbed Periodic
Fields. Phys. Rev. 97 (1955) 869–883
11. Morrow, R.A., Brownstein, K.R.: Model effective-mass Hamiltonians for abrupt
heterojunctions and the associated wave-function-matching conditions. Phys. Rev.
B 30 (1984) 678–680
12. Trefethen, L.N. , Bau III, D.: Numerical Linear Algebra (SIAM Society for Indus-
trial & Applied Mathematics, Philadelphia 1997)
13. Trellakis, A., Galick, A. T., Pacelli, A., Ravaioli, U.: Iteration scheme for the solu-
tion of the two-dimensional Schrödinger-Poisson equations in quantum structures.
J. Applied Phys. 81 (1997) 7880-7884
14. Varga, R. S.: Matrix Iterative Analysis (Prentice-Hall, Englewood Cliffs 1962)
15. Yu, P.Y., Cardona, M.: Fundamentals of Semiconductors (Springer, Berlin Heidel-
berg New York 1996)