This action might not be possible to undo. Are you sure you want to continue?
consistent solution of SchrOcUngerPoisson equations using a
nonuniform mesh
iH. Tan, G. L. Snider, L. D. Chang, and E. L. Hu
Department of Electrical alld Computer Engineering, University of California, Santa Barbara,
CalzJornia 93106
(Received 30 Apri11990; accepted for publication 28 June 1990)
A selfconsistent, onedimensional solution of the Schrodinger and Poisson equations is
obtained using the finitedifference method with a nonuniform mesh size. The use of the proper
matrix transformation allows preservation of the symmetry of the discretized Schrodinger
equation, even with the use of a nonuniform mesh. size, therefore reducing the computation
time. This method is very efficient in finding eigenstates extending over relatively large spatial
areas without loss of accuracy. For confirmation ofthe accuracy of this method, a comparison
is made with the exactly calculated eigenstates of GaAsl AIGaAs rectangular wells. An
example of the solution of the conduction band and the electron density distribution of a
singleheterostructure GaAsl AIGaAs is also presented.
l. iNTRODUCTiON
Growth of highquality heterostructure wens, together
with lateral feature modulation either by highresolution
fabrication processes 1,2 or, more recently, by innovative
growth on tilted substrates"" are producing structures and
devices of low dimensionality whose device implications
may be far reaching. Full understanding of the optical and
transport properties of these structures requires the selfcon
sistent solution of both Poisson and Schrodinger equa
tions.
5

H
A conventional approach to the sol.ution of the Schro
dinger equation has been the finitedifference method
(FDM). Real space is divided into discrete mesh points and
the wave function is solved within those discrete spacings.
Solving the differential equation within each mesh spacing
results in a vector solution for t/J, and a matrix formulation of
the Schrodinger equation:
(1)
where A is the matrix operator and IL the energy eigenvalues.
The implementation of FDM usually makes use of uniform
mesh spacings so that A is real and symmetric. The mesh size
is determined by the opposing requirements of greater accu
racy on one hand (hence, smaller mesh size) and rapid com
putation times on the other hand (hence large mesh size) .
However, there are many cases where the wave function
rapidly varies within one region, and then changes slowly
over another region. For example, Fig. 1 (a) shows the wave
function at the GaAsl AIGaAs single heterojunction, with
an energy eigenvalue dose to the Fermi level. In such a case,
the ideal situation would be to use a variable mesh size:
smaller mesh in region I, and large spacing in region II. The
use of a variable mesh, however, requires careful treatment
at the juncture of two different mesh sizes and destroys the
symmetry of the matrix A, in turn making the eigenfunction
more difficult to compute.
9
In th.is paper, we propose a sim
ple matrix transformation that will preserve the symmetry of
the matrix wh.ile allowing variable mesh size. If the optimal
nonuniform mesh is used, this method will provide a compu
tationally efficient solution of the band profile and the elec
tron density distributed over a large spatial dimension.
The equations used and the iterative procedure for ob
taining selfconsistent Schrodinger and Poisson solutions is
described in Sec. II. Section HI describes the matrix trans
formation used to preserve the symmetry of the discretized
Schrodinger equation and the Newton method to solve the
Poisson equation. A heterojunction quantum well and the
single heterostructure with modulated doping are given as
exemplary solutions in Sec. IV. Finally, the paper is sum
marized in Sec. V.
II. BASIC EQUATIONS
The onedimensional, oneelectron Schrodinger equa
tion is
conduction ~
band ,
AiGaAs
mesh size
mesh point index
f2giun I i region II
I
GaAs
FI G. 1. (a) Band diagram of the singleheterojunction GaAsl AIGaAs and
its bounded state wave functions. (b) Discretization oftne potential using a
nonuniform mesh.
4071 J. Appl. Phys. 68 (8), 15 October 1990 00218979/90/20407106$03.00 @ 1990 American Institute of PhySiCS 4071
Downloaded 12 Feb 2011 to 118.107.214.133. Redistribution subject to AIP license or copyright; see http://jap.aip.org/about/rights_and_permissions
_ fl2 ¢;(x) + V(x)¢,(x) = Et/J(x) , (2)
2 dx m*(x) dx
where 'I/J is the wave function, E is the energy, Vis the poten
tial energy, Ii is Planck's constant divided by 217", and m* is
the effective ma.<;s. The onedimensional Poisson equation is
€ q[Np(x)n(x)], (3)
dx dx Eo
where Es is the dielectric constant, tjJ is the electrostatic po
tential, N D is the ionized donor concentration, and n is the
electron density distribution. To find the electron distribu
tion in the conduction band, one may set the potential energy
V to be equal to the conductionband energy. In a quantum
well of arbitrary potential energy profile, the potential ener
gy V is related to the electrostatic potential if; as follows:
V(x) = qif>(x)+IlE
c
(x), (4)
where AEc is the pseudopotential energy due to the band
offset at the heterointerface. The wave function tP(x) in Eq.
(2) and the electron density n(x) in Eq. (3) are related by
m
n(x) = L 'l/Jt(X)'l/Jk (x)nk'
(5)
where m is the number of bound states, and n
k
is the electron
occupation for each state. The electron concentration for
each state can be expressed by
m* f"" 1
n
k
= .' dE,
1TfI2 Ek 1 + erE  EF)ikT
(6)
where Ek is the eigenenergy.
We use an iteration procedure to obtain selfconsistent
solutions for Eqs. (2) and (3). Starting with a trial potential
Vex), the wave functions, and their corresponding eigenen
ergies, Ek can be used to calculate the electron density distri
bution n(x) using Egs. (5) and (6). The computed n(x)
and a given donor concentration N D (x) can be used to cal
culatetjJ(x) viaEq. (3). Thenewpotentiaienergy Vex) is
then obtained from Eq. (4). The subsequent iteration will
yield the final selfconsistent solutions for V(x) and n (x)
which satisfy certain error criteria.
iii. NUMERICAL METHOD
A. Formulation for the equation
In order to numerically solve the Schr6dinger equation,
we may discretize the differential equation (3) by using a
threepoint finite difference scheme as shown in Fig. 1 (b) :
f/z ( 2( tPi+ 1  t/Ji )
2 mitll2hi(hi +hi _ l )
_ 2('I/J;'l/J, \) )=).'l/Ji'
mi_ll2hi_l (hi + hi_I)
This may be cast in the form of a matrix equation,
n
L AijtPj =A'l/Ji'
(7)
j= i
4072 J. Appl. Phys., Vol. 68, No.8, 15 October 1990
where
if j = i + 1,
if j = iI,
A
ii
+ 1 Aii. I + Vi
o
if j= i,
otherwise.
(8)
The index i identifies the grid point on the onedimensional
mesh. Halfinteger index implies a point midway between
the grid points, and hi is the mesh size between adjacent grid
points Xi and x
i
+ I' This gives a tridiagonal matrix which
appears symmetric only if the mesh spacings hi are all uni
form. While the use of a nonuniform mesh size might be
preferable for certain problems, this would destroy the sym
metry of the matrix A and hence obviate computational sim
plifications that result from the symmetry. However, if we
define the following parameter:
L f = (h; + hi_I )/2, (9)
Eq. (8) becomes
f/2 ( 1 _1_ if j = i + 1,
2 mt+ 1/2 hi L;
_ If ( 1 _1_) _1_ if j = iI,
2 m1 1/2 hi I L
o
if j= i,
otherwise.
We set Bij = L tAij' or in matrix notation,
B=MA,
(8')
(10)
where M is the diagonal matrix whose elements are L ;. As
the product of a diagonal and triadiagonal matix, B is tridia
gonal. From Eqs. (9) and (8'), it is easily seen that B is
symmetric, i.e., Bu; 1 = Bi t Ii' This provides the desired
transformation that allows us to solve for
B'i' = MA 1/J = )'M\II. (11 )
The matrix M obtained from the FDM method is diagonal so
that we may easily express it in the form
M=LL, (12)
where L is a diagonal matrix whose elements are Lt.
Using (10), we can show that
L  IEL lLt/J = L . ILLA'I/J =).L  lLL'I/J,
or
HI'P=AI'P, (13)
where
H=L lBL I
(14)
and
(15)
Since the relation H = H Tholds because B is symmetric and
L is diagonal, the matrix H is a symmetric and tridiagonal
matrix. Equation ( 13) is now the central equation for find
ing the eigenvalues and eigenfunctions. Therefore, instead of
Tan etal. 4072
Downloaded 12 Feb 2011 to 118.107.214.133. Redistribution subject to AIP license or copyright; see http://jap.aip.org/about/rights_and_permissions
solving Eq. (1), one may solve Eg. (13) to obtain the eigen
valueA corresponding to the eigenfunction <P. Then one may
use Eq. (IS) to obtain the wave function ¢from <1>. If!/J" and
¢I are two wave functions in the potential wen, they are
related to each other by
T {I ifk=l,
tPkMif;l = 0
if k # 1,
(16)
which means those wave functions are orthonormal to one
another with respect to the weighting matrix M.
It is noted that this method can be applied not only in
onedimensional cases but also in twodimensional ones. U s
ing the same approach to discretize the twodimensional
Schrodinger equation, we may obtain the same equation as
( 11) in which B turns out to be banded and symmetric and
J.f is still diagonal. Therefore, after the same matrix transfor
mation, we may obtain Eq. (13).
B. Formulation for Poisson's equation
Newton's method is used in solving the nonlinear Pois
son's equation (see the Appendix). The differential equation
describing the incremental potential change Or/> at each iter
ation satisfies the equation
 difJ)+ q [ND(x) n(x
n
]
dx dx Eo
= (€s d8¢) + ± tf/tt/Jk an
k
(t/Jdqo¢l¢k)'
dx \ dx Eo k 1 BE"
(17)
In the equation above, we have assumed that the donors are
completely ionized and the variation of wave function versus
8¢ is very smaiL The term on the lefthand side of the above
equation is the error and it is zero if the electrostatic poten
tial ¢ is equal to the selfconsistent solution. Although this
differentialintegral equation is hard to solve, a firstorder
approximation can be made by setting the integral result of
the braket to be q8¢. \0 From Eq. (6) we may also calculate
the partial derivative, an k I aE k' Therefore, using the same
method to discretize the Poisson equation as that used for
the Schrodinger equation, Eq. (17) can be expressed as
 ej = C
u
+ ll)¢H 1
where
"
ej = I Cij¢;j + q(N
Di
 nj )/"0,
j= I
2"i + 112
2Ei + 1/2
o
ifj = i + 1,
ifj = iI,
ifi =j,
otherwise.
(18)
(19)
(20)
Thus, in Eq. (18) there are n error equations with n un
4073 J. Appl. Phys., Vol. 68, NO.8, 15 October 1990
known variables (O¢l ,8¢2 , ... ,8¢" ). They may be manipUlat
ed as a matrix equation:
C'o¢ =  5, (21)
where C' is a tridiagonal, nonsymmetric, and n X n matrix,
/3¢; is the n X 1 vector containing the corrected potential at
each point which must be added to the former potential pro
files, and t is a n X 1 vector including the Poisson error at
each point. Now, Eq. (21) is the central equation required to
modify the former potential profile. To maximize the com
puting efficiency and to save the memory space, Crout's re
duction method!l is used for solving Eq. (21) where the
matrix C' is tridiagonal.
IV. RESULTS
A. Rectangular quantum weU
To evaluate the validity and the accuracy of the present
method using the nonuniform mesh, we first consider the
eigenstates of an electron in a simple rectangular quantum
weB of GaAs/Alo 3 Ga
o
.
7
As, as shown in Fig. 2(a). The
width of the well is chosen to be 56 A so that the upper (of
two) bound energy eigenvalues is very close to the height of
the potential barrier, so that the corresponding wave func
tion extends over a large distance outside the welt We con
sider the conduction band only so that the exact solutions
t potential
I energy
__________ ______ x
AIGaAs iGlIAs
AIGaAs
(lill
0.15
0.1
0.15 '__ "___ ...<..., ___ '___ 1
360 180 0 180
X position fA)
360
FIG. 2.(a) Rectangular 4uantum well of GaAs/AIGaAs:m*(GaAs)
= 0.0671110' m*(AIGaAsl =0 O.092mo, b.E, ''''' 0.23 eV. The variable mesh
size is chosen in different regions, ranging from 2 A inside the well to 32 A
far away from the well. (b) The bound eigenfunctions for the rectangular
quantum well.
Tan etal. 4073
Downloaded 12 Feb 2011 to 118.107.214.133. Redistribution subject to AIP license or copyright; see http://jap.aip.org/about/rights_and_permissions
TABLE I. Eigenenergies of GaAs/ Alo, Gao.7 As rectangular quantum well.
Uniform mesh (4 A) Uniform mesh (8 A) Nonuniform mesh (232 ft.)
Exact
n solution No. of FDM solution/error No. of FDM solution/error No. of FDM solution/error
(meV) meshes (meV/%) meshes (meV/%) meshes (meV/%)
64.24722 194 63.81747/0.669
2 220.7776 194 220.3347/0.201
can be obtained by solving the transcendental equations and
we can check the validity of our method. The transcendental
equations can be obtained from the connection rules (the
wave function and its derivative divided by the effective mass
continuous at the heterojunction of GaAsl AIGaAs. The
two bound wave functions are depicted in Fig. 2(b). The
width of the total region of interest is 768 A, 13 times larger
than that of the quantum well, to ensure the validity of the
boundary conditions.
Calculations using a nonuniform mesh (size ranging
from 2 to 32 A) have been made for the same potential. The
more standard computation, using a uniform mesh of 4 or 8
A size, was also carried out. Table I shows the calculated
eigenenergies of the quantum well, using both uniform mesh
and nonuniform mesh. The corresponding errors of the ei
genenergies, compared with the exact solutions are also list
ed. We observe that the percent error becomes smaller as the
the mesh size is reduced. The number of grids in the case of
nonuniform mesh is close to that in the case of uniform mesh
with 8 A size while the percent error of the former case is
much smaller than the latter one. We also notice that, in all
cases, the percent error for the second bound state is smaller
than that for the first bound state. This is because the second
bound state changes far less rapidly than the ground state, as
shown in Fig. 4.
B. Singleheterojunction GaAs/ AIGaAs
Now, consider a modulationdoped
GaAsl Al
a
.
3
Ga
o
.
7
As structure as shown in Fig. 3. In this
case, the electrons are weakly confined at the quasionedi
mensional heterojunction. The wave functions rapidly vary
within the kink of the conduction band, and then change
slowly over a long distance because of the smooth and slow
change of potential barrier.
A nonuniform mesh, ranging from 2 A at the vicinity of
the heterointerface to 32 A when the mesh is far away from
the heterointerface, is used in implementing this method. To
solve Poisson's equation, the boundary conditions imposed
on the structure are that the conduction band must equal the
barrier height at the surface, and the normal derivative must
be zero at the substrate. Likewise, the Neumanntype
boundary conditions have been assumed for the Schrodinger
equation, Since only bound states are considered, both the
wave function and its derivative should go to zero at the
boundary. If the wave function at the edge ofthe mesh is not
close to zero, the state is not a valid bound state. These
boundary conditions allow a quick check of the validity of
4074 J. Appl. Phys., Vol. 68, No.6, 15 October 1990
97
97
62.56356/2.621 104 64.28038/0.052
218.9136/0.844 104 220.6846/0.042
both wave functions and conduction band. For convenience,
the Fermi level in the conduction band has been chosen to be
the reference zero point. The temperature is assumed to be 4
K. To fulfill the requirements of the boundary conditions,
the width of the region is 4500 A large and the number of
mesh points is 450. The convergence criterion is that the
minimum of o<p, should be smaller than 10  5 V at the last
run of iteration. After the convergence criterion is satisfied, a
final check of the validity of the approximation using New
ton's method is made by comparing the ratio of the right
hand term to the lefthand term in Eq. (3). In general, the
difference between this ratio and 1 is smaller than 0.002 at all
mesh points, indicating that the approximation using New
ton's method is a good one. Once the solution for the discre
tized Poisson's equation has been found, it should be the
unique and selfconsistent solution for both equations. 12
There are two bound states in this case. After iterating
between the Schrodinger equation and the Poisson equation,
the convergent solutions of both conduction band and elec
tron density distribution are shown in Fig. 4. We observe
that the distance between the position of the peak electron
density and the heterointerface is 85 A. The tail of the elec
tron density distribution for x> 800 A primarily results
from the electrons occupying the second bound state.
v. CONCLUSION
In this paper, we present a finitedifference method to
selfconsistently solve the Schrodinger equation and the
Poisson equation. We use a simple matrix transformation
150A
GaAs
N =
D
1 X1 0 1 8
r
y
x
200;\
18
AIGaAs
No'"
1X1O
soA AIGaAs NiD
scooA
GaAs NID
GaAs substrate
PIG. 3. Structure of quasionedimensional channel in the modulation
doped GaAs/ Alo . .I Gao 7 As.
Tan et al. 4074
Downloaded 12 Feb 2011 to 118.107.214.133. Redistribution subject to AIP license or copyright; see http://jap.aip.org/about/rights_and_permissions
0.8
.
> 0.6
Q.I
'\
''
I
"C
' \
=
I .
,.Q
0.4
,
\
I
=
,
,
\
e
I
.

•
U
0.2
\
:I
'e
::
\
e
0
..
"
,
0.2
0 400 800 1200
x position
c
(A)
that allows use of nonuniform mesh size while preserving the
symmetric property of the Schrodinger matrix equation.
Newton's method is to solve the Poisson equation and find
the modified conduction band. This method is especially
suitable for finding the conduction band and the electron
density distributed over large spatial dimensions, with high
er computation efficiency and without loss of the accuracy.
Although we have only discussed the onedimensional
SchrodingerPoisson solver, it is straightforward to extend
this approach using a nonuniform mesh to the twodimen
sional cases. We foresee that the computation efficiency of
this method will be much higher than those using the con
ventional finitedifference or finiteelement methods.
ACKNOWLEDGMENTS
The authors wish to thank JongChang Yi and Young
chul Chung for their valuable discussion on this work. This
work was supported by the National Science Foundation
under the QUEST Science and Technology Center.
APPENDIX: NEWTON'S METHOD IN SOLVING
POISSON'S EQUATION
We show the generalization of Newton's method to
solve Eq. (3). Since the electron density is determined by
solutions of the Schr6dinger equation which are in turn de
termined by the potential ¢(x), the electron density is actu
ally a functional of ¢ (x) via Eqs. (3), (4), and (5). Denot
ing this functional by n [c;b]' Poisson's equation (3) can be
written as
..!£.(c d¢)= q(NDn[c;bJ).
dx S dx Eo
(Al)
Let us denote the exact solution to Eq. (At) by ¢(O) (x). For
a given trial function ¢ (x), our task is to find the correction
function 8¢(x) so that
4075 J. Appl. Phys., Vol. 68, No.8, i5 October 1990

10
n
=
 i>.

.....
00
10
16
==
W
FIG. 4. Conduction band and the
"0
electron density distribution in the
::
singleheterojunction GaAsl Al
o
.,
e
Ga
o
.
7
As.

y
.
10
15
eJ
1600 2000
(A2)
Substituting (A2) into (AI), we obtain
d¢)= q(NDn[¢+8¢]) d8¢).
dx dx Eo dx dx
Defining
n[¢ + o¢] = n[c;b] + Bn[¢],
Eq. (A3) can be written as
 +.i.. (N
D
 n[¢])]
dx ax EI)
d ( dB¢;) q
=  €S    8n [¢].
dx dx Eo
(A3)
(A4)
(A5)
Note that the lefthand side ofEq. (AS) is the error in Pois
son's equation for the trial function ¢(x) which can be easily
calculated. Assuming that B¢; (x) is small, from Eqs. (5) and
(AS), on [¢] can be expressed as
m
8n[¢] = I [8(rptlbd nk + t.bt¢k 8nd, (A6)
k _ 1
where
and
D( tl'ttl'k) = tl't[¢ + 0<;6 Nd¢ + O¢]  [¢] ¢d¢]
(A7)
(A8)
Our numerical experience indicates that the first term on the
righthand side ofEq. (A6) is usually much smaller than the
second one. Dropping the first term and expressing (jn k in
terms of o¢, explicitly, usi.ng Eqs. (4)  ( 6), we obtain
Tan etal. 4075
Downloaded 12 Feb 2011 to 118.107.214.133. Redistribution subject to AIP license or copyright; see http://jap.aip.org/about/rights_and_permissions
(A9)
where (I) is the braket integral.
Equations (AS)CA9) are now the formula to solve for
¢i using the Newton method.
I T. Demel, D. Heitmann, P. Grambow, and K. Ploeg, Appl. Phys. Lett.
53,2176 (1988).
2T. Thornton, M. Pepper, H. Ahmed, D. Andrews, and G. J. Davies, Phys.
Rev. Lett. 56,1198 (1986).
4076 J. Appl. Phys., Vol. 68, No.8, 15 October 1990
J J. M. Gaines, P. M. Petroff, H. Kroemer, R. J. Simes, R. S. Geels, and J.
H. English, J. Vac. Sci. Techno!. B 6, 1378 (1988).
4T. Fukui and H. Saito, App!. Phys. Lett. 50, 824 (1987).
SM. Tomizawa, T. Furuta, K. Yokoyama, and A. Yoshii, IEEE Trans.
Electron Devices ED·36, 2380 (1989).
6D. Mui, M. Patil, and H. Morko<;, App!. Phys. Lett. 55,1223 (1989).
7 G. L. Snider, IH. Tan, and E. L. Hu, J. Appl. Phys. 68, 2849 (1990).
8 S. E. Laux and F. Stern, AppJ. Phys. Lett. 49, 91 (1986).
9 F. Szidarouszky and S. Yakowitz. Principles and Procedures a/Numerical
Analysis (Plenum, New York, 1978), p. 221.
IOC. M. Krowne, J. App!. Phys. 65, 1602 (1989).
II J. Y.F. Tang and S. E. Laux, IEEE Trans. ComputerAided Design
CAD5, 645 (1986).
12 F. B. Hildebrand, Introduction to Numerical Analysis, 2nd ed. (Dover,
New York, 1987), p. 559.
Tan eta!. 4076
Downloaded 12 Feb 2011 to 118.107.214.133. Redistribution subject to AIP license or copyright; see http://jap.aip.org/about/rights_and_permissions
The computed n(x) and a given donor concentration N D (x) can be used to calculatetjJ(x) viaEq. (2) and (3). From Eqs. Equation ( 13) is now the central equation for finding the eigenvalues and eigenfunctions. (11 ) The matrix M obtained from the FDM method is diagonal so that we may easily express it in the form M=LL. we can show that L . (6) EF)ikT 2 m1 1 _1_) 1/2 hi L.107. (12) iii. o B=MA. ~(€'(X)~)tjJ(X)= dx q[Np(x)n(x)]. While the use of a nonuniform mesh size might be preferable for certain problems. if we define the following parameter: L = L k~1 m 'l/Jt(X)'l/Jk (x)nk' (5) f= (h. B is tridiagonal.org/about/rights_and_permissions . (8) becomes f/2 ( 2 where m is the number of bound states. where H=L lBL I 2( tPi+ 1  t/Ji ) l ) 2 mitll2hi(hi +h i _ (14) _ 2('I/J. and n is the electron density distribution. Vis the potential energy.s. = iI. (8) where Es is the dielectric constant.'l/J. i. this would destroy the symmetry of the matrix A and hence obviate computational simplifications that result from the symmetry. 1 (b) : f/z ( where L is a diagonal matrix whose elements are Lt. the wave functions. (4).. it is easily seen that B is symmetric. I + Vi if j= i. as follows: V(x) o = qif>(x)+IlEc (x). Ek can be used to calculate the electron density distribution n(x) using Egs. the matrix H is a symmetric and tridiagonal matrix. (8') m* nk = 1TfI2 f"" Ek 1 + erE  1 . see http://jap. In a quantum well of arbitrary potential energy profile. and hi is the mesh size between adjacent grid points Xi and x i + I ' This gives a tridiagonal matrix which appears symmetric only if the mesh spacings hi are all uniform.. j= i L AijtPj =A'l/Ji' J. 1 = Bi t. Using (10). and n k is the electron occupation for each state. Ii is Planck's constant divided by 217".IEL lLt/J = L .L . ILLA'I/J =).' _ If ( dE. one may set the potential energy V to be equal to the conductionband energy. 68. Therefore.214. We use an iteration procedure to obtain selfconsistent solutions for Eqs.. (2) and the electron density n(x) in Eq. Halfinteger index implies a point midway between the grid points. (5) and (6).aip. 1 (9) Eq. 15 October 1990 n (7) Since the relation H = H Tholds because B is symmetric and L is diagonal.8. (3) are related by n(x) The index i identifies the grid point on the onedimensional mesh. \) )=). if j= i.e.(x) = Et/J(x) . _1_ L ~ hi I where Ek is the eigenenergy. Starting with a trial potential Vex). As the product of a diagonal and triadiagonal matix. The subsequent iteration will yield the final selfconsistent solutions for V(x) and n (x) which satisfy certain error criteria. The electron concentration for each state can be expressed by mt+ ~) _1_ 1/2 if j if j = i + 1. and their corresponding eigenenergies. otherwise.(x) + V(x)¢. (3). Thenewpotentiaienergy Vex) is then obtained from Eq. Vol. However. Eo (3) dx A ii + 1 Aii.<. Phys. The onedimensional Poisson equation is if j = i + 1.Ii' This provides the desired transformation that allows us to solve for B'i' = MA 1/J = )'M\II. we may discretize the differential equation (3) by using a threepoint finite difference scheme as shown in Fig. otherwise.lLL'I/J. + hi_I )/2. the potential energy V is related to the electrostatic potential if.133. tjJ is the electrostatic potential. 4072 4072 Downloaded 12 Feb 2011 to 118. Formulation for the Schr~dinger equation In order to numerically solve the Schr6dinger equation._ fl2 ~ (_1_~) ¢. instead of Tan etal. (13) or HI'P=AI'P. N D is the ionized donor concentration.'l/Ji' mi_ll2hi_l (hi + hi_I) and (15) This may be cast in the form of a matrix equation. No. E is the energy. To find the electron distribution in the conduction band. (2) where 2 dx m*(x) dx where 'I/J is the wave function. (10) We set Bij = L tAij' or in matrix notation. and m* is the effective ma. (4) where AEc is the pseudopotential energy due to the band offset at the heterointerface. Redistribution subject to AIP license or copyright. NUMERICAL METHOD A. Bu. (9) and (8'). where M is the diagonal matrix whose elements are L . Appl. The wave function tP(x) in Eq. if j = iI.
The variable mesh size is chosen in different regions. (21) tPkMif..j + q(N 2"i + Di  nj )/"0. so that the corresponding wave function extends over a large distance outside the welt We consider the conduction band only so that the exact solutions ~ (€s dx \ d8¢) dx +~ Eo tf/tt/Jk an (t/Jdqo¢l¢k)' k~ 1 BE" (17) ± In the equation above.0671110' m*(AIGaAsl O._ _ _. ranging from 2 Ainside the well to 32 A far away from the well. ifi =j. NO.aip. is the n X 1 vector containing the corrected potential at each point which must be added to the former potential profiles. and is a n X 1 vector including the Poisson error at each point. which means those wave functions are orthonormal to one another with respect to the weighting matrix M.ej = C u + ll)¢H 1 t I potential energy __________kI~I~I_o... =0 Tan etal.. Then one may use Eq. we may obtain the same equation as ( 11) in which B turns out to be banded and symmetric and J. (b) The bound eigenfunctions for the rectangular quantum well. using the same method to discretize the Poisson equation as that used for the Schrodinger equation.7 As. (19) 0. Eq. they are related to each other by ifk=l. . after the same matrix transformation. The differential equation describing the incremental potential change Or/> at each iteration satisfies the equation . ''''' 0.[~(Es difJ)+ q dx dx = [ND(x) n(x Eo n] k To evaluate the validity and the accuracy of the present method using the nonuniform mesh.. Redistribution subject to AIP license or copyright.8¢2 . Therefore. They may be manipUlated as a matrix equation: C'o¢ =  5. (IS) to obtain the wave function ¢from <1>.1 (18) where ej = I" j= I Cij¢. (6) we may also calculate the partial derivative. a firstorder approximation can be made by setting the integral result of the braket to be q8¢. \0 From Eq.107. known variables (O¢l . 2(a). (18) there are n error equations with n un4073 J.8. (21) is the central equation required to modify the former potential profile. Thus. If!/J" and ¢I are two wave functions in the potential wen. /3¢. (21) where the matrix C' is tridiagonal.solving Eq.8¢" ). RESULTS A. 68. Phys.214.23 eV. (1). Formulation for Poisson's equation Newton's method is used in solving the nonlinear Poisson's equation (see the Appendix).f is still diagonal. see http://jap. Now. an k I aEk' Therefore. and n X n matrix. (17) can be expressed as . (13) to obtain the eigenvalueA corresponding to the eigenfunction <P.. otherwise.15 '_ _"_ _ _.15 0. t IV. we have assumed that the donors are completely ionized and the variation of wave function versus 8¢ is very smaiL The term on the lefthand side of the above equation is the error and it is zero if the electrostatic potential ¢ is equal to the selfconsistent solution. b.092mo. Crout's reduction method!l is used for solving Eq.org/about/rights_and_permissions . one may solve Eg. in Eq.' _ _ _1 360 112 ifj = i + 1. 4073 Downloaded 12 Feb 2011 to 118. nonsymmetric. Although this differentialintegral equation is hard to solve. The width of the well is chosen to be 56 A so that the upper (of two) bound energy eigenvalues is very close to the height of the potential barrier... (13). as shown in Fig. Rectangular quantum weU B. Appl. (16) if k # 1. we first consider the eigenstates of an electron in a simple rectangular quantum weB of GaAs/Alo 3 Gao. It is noted that this method can be applied not only in onedimensional cases but also in twodimensional ones. 2. (20) 180 X 0 position fA) 180 360 2Ei + 1/2 ifj = iI.E.. U sing the same approach to discretize the twodimensional Schrodinger equation.133. Eq. Vol. we may obtain Eq. To maximize the computing efficiency and to save the memory space. 15 October 1990 o FIG.l T = {I0 where C' is a tridiagonal.<.(a) Rectangular 4uantum well of GaAs/AIGaAs:m*(GaAs) = 0.2~:_e_v______+~ AIGaAs x iGlIAs AIGaAs (lill 0.
3 Gao. Once the solution for the discretized Poisson's equation has been found. The wave functions rapidly vary within the kink of the conduction band.052 220. 3.107.3347/0. of meshes A) Nonuniform mesh (232 ft.81747/0. After the convergence criterion is satisfied. the width of the region is 4500 A large and the number of mesh points is 450. the state is not a valid bound state.28038/0. The two bound wave functions are depicted in Fig. as shown in Fig. 4. No.7776 No. The width of the total region of interest is 768 A. 13 times larger than that of the quantum well. 4074 Downloaded 12 Feb 2011 to 118. the percent error for the second bound state is smaller than that for the first bound state. The tail of the electron density distribution for x> 800 A primarily results from the electrons occupying the second bound state. We observe that the distance between the position of the peak electron density and the heterointerface is 85 A. the Fermi level in the conduction band has been chosen to be the reference zero point. was also carried out. both the wave function and its derivative should go to zero at the boundary.042 63.. The transcendental equations can be obtained from the connection rules (the wave function and its derivative divided by the effective mass continuous at the heterojunction of GaAslAIGaAs. Singleheterojunction GaAs/ AIGaAs Now. We also notice that. CONCLUSION B. Appl. Likewise. consider a modulationdoped GaAslAl a.TABLE I. in all cases. 15 October 1990 In this paper. In general. to ensure the validity of the boundary conditions.. using both uniform mesh and nonuniform mesh. Phys. If the wave function at the edge ofthe mesh is not close to zero.669 220. a final check of the validity of the approximation using Newton's method is made by comparing the ratio of the righthand term to the lefthand term in Eq. the boundary conditions imposed on the structure are that the conduction band must equal the barrier height at the surface.aip. We use a simple matrix transformation 150A 200.56356/2.844 FDM solution/error (meV/%) 64. the Neumanntype boundary conditions have been assumed for the Schrodinger equation. To fulfill the requirements of the boundary conditions. 12 There are two bound states in this case.org/about/rights_and_permissions .7 As rectangular quantum well. These boundary conditions allow a quick check of the validity of 4074 J.5 V at the last run of iteration. (3).\ GaAs ND = 1 X1 0 1 8 18 r x y AIGaAs AIGaAs GaAs No'" 1X1O NiD NID soA scooA GaAs substrate PIG. The number of grids in the case of nonuniform mesh is close to that in the case of uniform mesh with 8 A size while the percent error of the former case is much smaller than the latter one.9136/0. both wave functions and conduction band.201 97 97 can be obtained by solving the transcendental equations and we can check the validity of our method.) No.6. of meshes 194 194 A) Uniform mesh (8 No. Table I shows the calculated eigenenergies of the quantum well. The temperature is assumed to be 4 K. Eigenenergies of GaAs/Alo. it should be the unique and selfconsistent solution for both equations.I Gao 7 As. The more standard computation. We observe that the percent error becomes smaller as the the mesh size is reduced. of meshes 104 104 FDM solution/error (meV/%) FDM solution/error (meV/%) 62. 4.7 As structure as shown in Fig. In this case.133. and the normal derivative must be zero at the substrate.621 218. compared with the exact solutions are also listed. To solve Poisson's equation. and then change slowly over a long distance because of the smooth and slow change of potential barrier.6846/0. Structure of quasionedimensional channel in the modulationdoped GaAs/ Alo . Since only bound states are considered. should be smaller than 10 . For convenience. the electrons are weakly confined at the quasionedimensional heterojunction. is used in implementing this method. Redistribution subject to AIP license or copyright. This is because the second bound state changes far less rapidly than the ground state. see http://jap. using a uniform mesh of 4 or 8 A size. A nonuniform mesh. the convergent solutions of both conduction band and electron density distribution are shown in Fig. Uniform mesh (4 n Exact solution (meV) 64. Calculations using a nonuniform mesh (size ranging from 2 to 32 A) have been made for the same potential.24722 2 220. Vol. 2(b). Gao. After iterating between the Schrodinger equation and the Poisson equation. the difference between this ratio and 1 is smaller than 0. v. ranging from 2 A at the vicinity of the heterointerface to 32 A when the mesh is far away from the heterointerface.214. The convergence criterion is that the minimum of o<p. The corresponding errors of the eigenenergies. 68. 3. we present a finitedifference method to selfconsistently solve the Schrodinger equation and the Poisson equation. indicating that the approximation using Newton's method is a good one.002 at all mesh points. Tan et al.
[¢] can be expressed as on We show the generalization of Newton's method to solve Eq. see http://jap.107. dx ax EI) (ND  n[¢])] (A5) The authors wish to thank JongChang Yi and Youngchul Chung for their valuable discussion on this work. (x) is small. 00 "0 0 .bt¢k 8n d.b]' Poisson's equation (3) can be written as 8n[¢] where = k _ 1 I m [8(rptlbd n k + t. (A3) can be written as . Denoting this functional by n [c.4 :I 'e ~ U 0.8 .Q ~ "C 0. For a given trial function ¢ (x). Dropping the first term and expressing (jn k in terms of o¢. dx Eo (Al) Let us denote the exact solution to Eq.. (5) and (AS). explicitly.ng Eqs. 4. No..2 :: e I ..214. with higher computation efficiency and without loss of the accuracy. ACKNOWLEDGMENTS (A2) Substituting (A2) into (AI). 1200 c 10 15 1600 2000 x position (A) that allows use of nonuniform mesh size while preserving the symmetric property of the Schrodinger matrix equation. Note that the lefthand side ofEq. (At) by ¢(O) (x).133.aip. This method is especially suitable for finding the conduction band and the electron density distributed over large spatial dimensions.t/. and (5). 4075 Downloaded 12 Feb 2011 to 118.i.6 = = e .~ \ I . (A4) Eq..bJ). from Eqs.6 Nd¢ + O¢] .d ( €S dB¢. the electron density is actually a functional of ¢ (x) via Eqs. \ I • \ \ ' '\ 10 n ~ = ~ ~ 10 16 .7 As.2 " . (4).( 6).[~(E. usi.. 0 400 800 ~ . Since the electron density is determined by solutions of the Schr6dinger equation which are in turn determined by the potential ¢(x). Assuming that B¢. .. Newton's method is to solve the Poisson equation and find the modified conduction band. > '' Q. 0. Although we have only discussed the onedimensional SchrodingerPoisson solver.) dx dx Eo q 8n [¢]. Conduction band and the electron density distribution in the singleheterojunction GaAsl Alo. Gao.~ [¢] ¢d¢] (A7) . our task is to find the correction function 8¢(x) so that 4075 J.. Appl. APPENDIX: NEWTON'S METHOD IN SOLVING POISSON'S EQUATION = . (3). Redistribution subject to AIP license or copyright. e y eJ == :: W FIG.I 0. dx Defining n[¢ + o¢] dx Eo dx dx (A3) = n[c. We foresee that the computation efficiency of this method will be much higher than those using the conventional finitedifference or finiteelement methods. (3). we obtain Tan etal.. (4) . i>..8.(c dx S and (A8) d¢)= q(NDn[c. ~¢) +. i5 October 1990 Our numerical experience indicates that the first term on the righthand side ofEq.org/about/rights_and_permissions . (AS) is the error in Poisson's equation for the trial function ¢(x) which can be easily calculated. 68.. we obtain ~(cs d¢)= q(NDn[¢+8¢]) ~(Es d8¢). it is straightforward to extend this approach using a nonuniform mesh to the twodimensional cases. \. This work was supported by the National Science Foundation under the QUEST Science and Technology Center. I .. Vol.b] + Bn[¢]. (A6) is usually much smaller than the second one.0. (A6) D( tl'ttl'k) = tl't[¢ + 0<.!£. Phys.
ComputerAided Design CAD5.1198 (1986). P. Andrews. 50. Davies. Heitmann. L. Appl. L.214. Saito. Gaines. Yokoyama.2176 (1988). App!. H. and G. 55. 7 G. Lett. Snider. Hu. Ahmed. and E. Petroff.F. 49. II J. Y. Hildebrand. Phys.(A9) where (I) is the braket integral. Phys. 645 (1986). and J. IEEE Trans. 68. M. Phys. 4076 Downloaded 12 Feb 2011 to 118. Lett. J. Lett. Phys. Appl. H.. R. Grambow. Furuta. Principles and Procedures a/Numerical Analysis (Plenum. IH. Geels. M. J. 824 (1987). Lett. p. Redistribution subject to AIP license or copyright.aip. Phys. New York. 91 (1986). 2nd ed. 56. 1602 (1989). Krowne. Demel. Fukui and H. Ploeg.8. J. 8 S. S. Patil. Rev. New York.1223 (1989). 2849 (1990). Laux and F. Mui. Phys. No. App!.107. Techno!. Tang and S. IOC. M. 1978). 53. Thornton. B. Kroemer. Tomizawa. Szidarouszky and S. K. 2380 (1989). D. Laux. I J. Tan. 68. 1378 (1988). see http://jap. P. D. Pepper. Electron Devices ED·36. and A. Introduction to Numerical Analysis. M. App!. 2T. R. T. Simes. M.. (Dover. IEEE Trans. Equations (AS)CA9) are now the formula to solve for ¢i using the Newton method. 65. Morko<. 6D. Stern. 12 F. Appl. Lett. Phys. T. H.133. Vol. 559. E. 221. p. J. 9 F. Sci. 15 October 1990 Tan eta!. B 6. English. 1987).org/about/rights_and_permissions . J. Yoshii. Phys. AppJ. J 4076 J. E. and H. Yakowitz. SM. 4T. and K. Vac.