A ballistic transport and surface reaction model for simulating atomic
layer deposition processes in highaspectratio nanopores
Raymond A. Adomaitis
∗
Department of Chemical and Biomolecular Engineering
Institute for Systems Research
University of Maryland
College Park, MD 20742 USA
Abstract
In this paper we develop a model describing the ballistic transport of chemical precursor species
for an atomic layer deposition process (ALD). In the application we consider, pore geometry or sur
face properties of nanoporous materials are modiﬁed using ALD, taking advantage of its potential for
conformal deposition in highaspectratio structures. Because of the very large Knudsen number corre
sponding to these processes, the transport of gasphase species inside the nanostructures takes place in
a purely ballistic manner. Precursor transmission probability functions describing the ﬂuxes between the
pore surface features are developed and compared to previously published results. The transport model
elements then are coupled to ALD surface reaction models, spatially discretized, and integrated over
each precursor exposure period to determine the pore spatial surface reaction extent proﬁle. Predictions
from our dynamic model then are compared to four previously published studies of ALD in nanopores
to validate our simulator and to gain further insight into the physical mechanisms at work in ALD
processes. The utility of physically based models of the type we develop can be exploited to determine
optimal precursor exposure levels for ALD based nanomanufacturing operations.
1 Introduction
Atomic layer deposition (ALD) is a thinﬁlm manufacturing process in which the growth surface is
exposed to an alternating sequence of gasphase chemical precursor species. The selflimiting nature of
the precursor chemisorption in ideal ALD processes makes possible highly conformal surface coverage
and atomiclevel control of ﬁlm thickness. While conceptually simple, ALD processes present a number
of true modeling and process optimization challenges. These diﬃculties stem from the wide ranging
time and length scales associated with the growing ﬁlm, the modeling challenges posed by transport of
∗
Corresponding author: adomaiti@umd.edu
2
precursor and product species through the reactors, and the complexity of the ﬁlm growth mechanisms.
Quantiﬁcation of the timescales associated with ALD surface reactions and micro/reactorscale
transport of precursor species makes it possible to determine the minimum exposure level necessary to
reach the selflimited regime. Most current processes and those under development for new industrial
applications compensate for nonideal reactor designs by signiﬁcantly overexposing the growth surface
to precursors, resulting in low process gas conversion rates. Because of promising ALD applications
in largescale micro and nanodevice fabrication and the potential for rapid expansion of this thin
ﬁlm manufacturing technology, we present a modeling approach suited to describing the dynamics of
precursor transport when coupled to ALD growth models. In this paper, we examine the complexities of
numerically solving these modeling equations and compare our results to previously published simulation
and experimental studies.
1.1 Previous work
The importance of ALD timescale analysis from an industrial perspective is presented in Granneman
et al. [11] where several conﬁgurations of singlewafer reactors and multiwafer batch reactor systems
are analyzed. In the cited study, the authors break down the timescale analysis to ALD precursor
chemisorption on ﬂat substrates and within highaspectratio cylindrical pores; these processes then
are combined with overall precursor material balances corresponding to the macroscopicscale transport
through each reactor system. The ﬂat surface and pore saturation rates are computed using the simpliﬁed
models of Gordon et al. [10]. In Gordon’s work, surface adsorption site material balances are combined
with precursor ﬂux calculations to yield the minimum exposure (measured in pressure × time) necessary
to saturate a ﬂat ALD growth surface. These computations were extended to ALD in pores, where
the precursor ﬂux was modulated by a Clausing factor to account for the reduced ﬂux within the pore.
Gordon’s model was enhanced by Dendooven and coworkers [5] who included a Langmuirtype surface
coverage model for the ALD reaction, demonstrating improved prediction of the ﬁlm thickness proﬁles
inside pores for which exposure is insuﬃcient to completely saturate the pore walls. Alumina ALD in
nanopores of extremely high aspect ratio were investigated by Elam et al. [6] using a onedimensional
Monte Carlo simulation procedure. Likewise, Knoops et al. [12] developed a 2dimensional Monte
Carlo simulation of ALD in highaspectratio trenches to investigate the deposition characteristics of
plasmaassisted ALD, examining the eﬀects of a range of sticking and radical species recombination
3
coeﬃcients.
Because of the very large Knudsen number corresponding to lowpressure precursors and the nanoscale
diameter of the porous structures under consideration [1], the transport of gasphase species inside
nanostructures takes place in a purely ballistic manner. In a pioneering sequence of lowpressure CVD
[2, 3] and ALD [9] theoretical modeling papers produced by Cale and coworkers, rigorous ballistic trans
port models for steady deposition processes and direct solution of the Boltzmann’s equations describing
molecularlevel precursor transport dynamics were developed. Cale’s approach to combining the ballistic
transport material balances with surface reaction rates was extended to reactive ion etching processes
to compute neutral species spatial ﬂux proﬁles [13] and the evolution of the etch front [18], phenomena
that have strong parallels to modeling ALD in porous structures.
In this paper, we will derive expressions for precursor transmission probability functions describing
the ﬂuxes between the pore surface features based on Cale’s work. These transport models are coupled
to ALD surface reaction models, spatially discretized, and integrated over each precursor exposure period
to determine the pore spatial surface reaction extent proﬁle. Models of this form will be shown eﬀective
in predicting experimentally observed deposition proﬁles in highaspectratio pores, giving a picture of
the ALD process dynamics during each exposure phase.
m
e
m
b
r
a
n
e
s
u
r
f
a
c
e
AAO substrate
2R
z=0
z=L
pore mouth
pore bottom
n
dA
z
!
dA’(r’,!’)
s
n'
"
z = 0
"' x
y
! = 0
Figure 1: A pore with AR=(L/2R)=10, left end open to precursors and closed at right by the pore
bottom (left ﬁgure). Diﬀerential area elements located at the pore entrance z
and a point within the
pore, z; precursor material balance modeling takes place at z (right ﬁgure).
2 Intrapore ballistic transport model
Consider precursor molecules diﬀusing into a cylindrical pore with major axis perpendicular to the
substrate surface. The geometry of the pore is shown in Fig. 1 where the pore entrance is located at
4
the axial coordinate position z = z
= 0. We assume the pore has a uniform (with respect to its axial
coordinate z) radius R.
2.1 Pore entrance eﬀects
Deﬁning a diﬀerential element at the pore entrance dA
= r
dr
dφ
located at z = z
= 0 and
(r
, φ
), the ﬂux of precursor species normal to a hemisphere of radius s is
η
e
(r
, φ
, s) =
η
o
cos Ω
dA
2πs
2
c
0
where η
o
is the idealgas ﬂux at the pore entrance plane (given by equation (6) deﬁned later in this paper)
and c
0
= 1/2 guarantees species conservation. We assume the precursor velocity direction component is
isotropically distributed at and above the pore opening. The vector s connects the entrance diﬀerential
element dA
to an element located at (z, r = R, φ = 0) on the inner wall of the pore in Fig. 1—choosing
φ = 0 results in no loss of generality in the following analysis. Therefore, the total ﬂux at the pore wall
element dA = Rdφdz contributed by the direct ﬂight of precursor molecules passing through the pore
entrance at z = 0 (i.e., not the ﬂux resulting from the reemission of precursor species from the pore
wall) is
η
ew
(z) = η
o
_
R
0
_
2π
0
cos Ω
cos Ω
πs
2
dA
an expression consistent with that of [2]. Based on the pore geometry
n = y
n
= z
s = (0 −r
sin φ
)x + (−R + r
cos φ
)y + (z −0)z
s
2
= (r
)
2
+ R
2
−2Rr
cos φ
+ z
2
and so
cos Ω
=
n
· s
n
 s
=
z
_
(r
)
2
+ R
2
−2Rr
cos φ
+ z
2
and
cos Ω =
−n · s
n s
=
R −r
cos φ
_
(r
)
2
+ R
2
−2Rr
cos φ
+ z
2
.
5
0 50 100 150
0
0.1
0.2
0.3
0.4
0.5
0.6
z
Q
e
w
0 10 20 30
0.036
0.0365
0.037
0.0375
0.038
r
Q
e
b
0 50 100 150
2
4
6
8
10
12
14
16
x 10
ï3
zv
Q
w
w
z=0
0 50 100 150
0
0.005
0.01
0.015
0.02
0.025
0.03
zv
Q
w
b
r=0
0 10 20 30
2
4
6
8
10
12
14
x 10
ï4
rv
Q
b
w
z=0
Figure 2: Transmission probability proﬁles for R = 30 nm and L = 150 nm illustrating Q
ew
(z), Q
eb
(r ),
Q
ww
(z, z
), Q
wb
(r , z
), and Q
bw
(z, r
). For those transmission functions used in a convolution, the color
change (from blue to red) indicates a change in spatial position from minimum to maximum value.
6
The ﬁnal form of the incident ﬂux (molecules/area/time) along the pore wall resulting from the precursors
diﬀusing through the pore entrance is
η
ew
(z) =
_
R
0
_
2π
0
η
o
z(R −r
cos φ
)
π [(r
)
2
+ R
2
−2Rr
cos φ
+ z
2
]
2
dφ
r
dr
= η
o
Q
ew
(z) =
η
o
2R
_
z
2
+ 2R
2
√
z
2
+ 4R
2
−z
_
.
Using the notation of Cale [2], Q
ew
(z) denotes the dimensionless transmission probability of a precursor
species traveling through the entire pore entrance to dA. A proﬁle of Q
ew
(z) is shown in Fig. 2.
2.2 Summary of transmission probabilities
For interactions between the inlet and bottom, the wall and itself, and the wall and the pore bottom,
derivation details are provided in the Appendix. A listing of all pore entrancesurface and surfacesurface
ﬂuxes in terms of each transmission probability function is as follows:
EntranceWall η
ew
(z) = η
o
Q
ew
(z)
EntranceBottom η
eb
(r ) = η
o
Q
eb
(r )
WallWall η
ww
(z) =
_
L
0
w
(z
)Q
ww
(z, z
) dz
WallBottom η
wb
(r ) =
_
L
0
w
(z
)Q
wb
(r , z
) dz
BottomWall η
bw
(z) =
_
R
0
b
(r
)Q
bw
(z, r
) r
dr
where
w
(z
) and
b
(r
) are the precursor reemission ﬂuxes from the pore wall and bottom, respectively.
The transmission probability functions themselves are listed below and are plotted in Fig. 2.
Q
ew
(z) =
1
2R
_
z
2
+ 2R
2
√
z
2
+ 4R
2
−z
_
Q
eb
(r ) =
1
2
_
_
1 −
L
2
+ r
2
−R
2
_
(L
2
+ r
2
−R
2
)
2
+ 4L
2
R
2
_
_
Q
ww
(z, z
) =
1
4R
_
_
2 −
3z −z

_
(z −z
)
2
+ 4R
2
+
z −z

3
((z −z
)
2
+ 4R
2
)
3/2
_
_
Q
wb
(r , z
) =
2R
2
(L −z
)
_
(L −z
)
2
+ R
2
−r
2
¸
_
[(L −z
)
2
+ R
2
+ r
2
]
2
−4R
2
r
2
_
3/2
Q
bw
(z, r
) =
2R(L −z)
_
(L −z)
2
+ R
2
−(r
)
2
¸
_
[(L −z)
2
+ R
2
+ (r
)
2
]
2
−4R
2
(r
)
2
_
3/2
7
The transmission probabilities listed above are nearly identical to those in Table I of Cale [2] with the
only diﬀerence being the constant resulting from integration over the pore circumference.
3 Comparison to previous Knudsen transport results
In an overview of rariﬁed gas ﬂow through tubes [17], Steckelmacher provides a comparative analysis
of the transmission probabilities computed by various researchers, including very precise calculations over
a range of L/R from 0.1−100 (in Table 3 of the cited work). We use these values to assess the accuracy
of our numerical techniques before moving on to the ALD simulation.
For tubes open at both ends with precursor supplied at z = 0 and with the other end at z = L
exposed to vacuum, a steadystate ﬂux balance gives
η
ew
(z) + η
ww
(z) −
w
(z) = 0
at each point along the tube wall. We note that no surface reaction or adsorption takes place under these
conditions. The two unknowns in the equation above are η
ww
(z) and
w
(z). Discretizing the spatially
dependent quantities using n
z
axial and n
r
radial collocation points and corresponding quadrature
weights w
z
i
and w
z
j
(the collocation details are described later in this manuscript), we can write the
complete set of equations as follows
η
ww
(z
i
) −
w
(z
i
) = −η
ew
(z
i
) (1)
η
ww
(z
i
) −
nz
k=1
w
(z
k
)Q
ww
(z
i
, z
k
)w
z
k
= 0 (2)
for i = 1, ... , n
z
with the entrance and exit ﬂuxes explicitly deﬁned by
η
ew
(z
i
) = η
o
Q
ew
(z
i
) (3)
η
wb
(r
j
) =
nz
k=1
w
(z
k
)Q
wb
(r
j
, z
k
)w
z
k
(4)
for j = 1, ... , n
r
. This set of linear equations can be written in matrix form
Ax = b with x =
_
¸
_
η
ww
w
_
¸
_ (5)
8
and x ∈ R
2nz
with net pore transmission rate
2π
πR
2
η
o
_
R
0
(η
eb
+ η
wb
) rdr =
2
R
2
η
o
nr
j =1
_
η
o
Q
eb
(r
j
)w
r
j
+
nz
k=1
w
(z
k
)Q
wb
(r
j
, z
k
)w
z
k
_
.
Solving (5) directly for x and determining the net pore transmission rate, the results of our modeling
procedure are illustrated in Fig. 3 where the solid curve corresponds to our model prediction and circles
denote data from [17].
0 20 40 60 80 100
0
0.2
0.4
0.6
0.8
1
L/R
p
o
r
e
t
r
a
n
s
m
i
s
s
i
o
n
p
r
o
b
a
b
i
l
i
t
y
Figure 3: Validation of our transport model (solid curve) by comparison to previously published results
[17] shown as circles.
3.1 Quadrature
There are two important aspects to solving (1)(4) that are nontrivial: the discontinuous derivative
of Q
ww
at z = z
and the lack of limits for Q
wb
at (r , z) = (R, L) discussed in the Appendix and in [13].
In the discretization procedure, the quadrature weights w
z
can be computed in a straightforward manner
by ﬁrst deﬁning the collocation points z
k
in the interval z ∈ [0, L] with n
z
≥ 2 as z
k
= Lk/(n
z
− 1),
k = 0, 1, ... , n
z
−1 and then using the trapezoidal rule to deﬁne the x
z
k
. This works [4], but converges
slowly, and we attribute the poor performance to the nature of Q
ww
(z, z
) which has a discontinuous
derivative at z = z
. Therefore, we consider representing the ﬂuxes and transmission probability functions
as continuous functions with piecewise continuous derivatives in subintervals of z ∈ [0, L] to accurately
integrate the walltowall ﬂux relationship; we use a multiplegrid technique to likewise improve the
9
accuracy with which the ﬂuxes and other transmission probabilities are represented.
To implement our discretization procedure, we develop a hierarchical quadrature grid architecture.
The most coarse and globally deﬁned level (r, z) is used to discretize ﬂux terms such as η
ww
(z), η
bw
(z),
η
wb
(r ), etc. Also deﬁned globally, but at a substantially ﬁner level are (¯r, ¯z) used to approximate the
more rapidly changing transmission probability functions such as Q
ew
(z) for highaspectratio pores.
Lastly, ﬁne grids over the appropriate subdomains in z are deﬁned for Q
ww
(z, z
). The interpolating
functions and arrays and the quadrature weights are derived in the Appendix.
4 Coupling pore transport to surface reactions
Despite the attention paid to even the most intensely studied ALD system – alumina ALD using
water and trimethylaluminum (TMA) as precursors [16] – much remains unknown regarding the speciﬁc
reactions that take place on the growth surface, and no consistent picture has yet emerged in terms of
the reaction kinetics of the overall ﬁlm growth process. As a result of the signiﬁcant experimental (e.g.,
[14]) and computational chemistry (e.g., density functional theory reaction modeling, [7, 19]) studies,
elements of the complete deposition reaction mechanism are beginning to emerge [8]. Most ALD surface
reaction kinetics modeling work relates surface dynamics to precursor impingement ﬂuxes computed using
kinetic theory of gases [10], frequently modiﬁed using a sticking coeﬃcient. Langmuirtype adsorption
mechanisms modiﬁed by a sticking coeﬃcient [5] also have been developed. Gobbert and coworkers
[9] assumed ALD reactions in the form of EleyRideal kinetics where precursor A adsorbed reversibly
followed by an irreversible reaction with B to form the solid ﬁlm. To date, Elliott [8] has produced
one of the most comprehensive frameworks for the elementary steps of ALD reactions, breaking down
the reaction process into chemisorption, dissociation of surface species, diﬀusion of ions on the growth
surface, and association of these ions to form desorption products.
We consider ALD reactions of the form
ν
A
A + ν
B
B →bulk ALD ﬁlm
where the stoichiometric coeﬃcients ν
i
, i = A,B indicate the number of atoms each precursor species
contributes to the ideal ALD ﬁlm. We take ν
A
= 1 for this study, and so for the case of alumina ALD
using TMA and water as precursors, we see that ν
B
= 3/2 by the composition of the minimal unit of
10
the alumina ﬁlm AlO
3/2
TMA +
3
2
H
2
O →AlO
3/2
+ 3CH
4
.
By the kinetic theory of gases, the precursor ﬂux at the ﬁlm growth surface at the precursor partial
pressure P
i
is
η
i
o
=
P
i
_
2πM
i
w
R
g
T
mol/(m
2
s) (6)
where M
i
w
is the molecular mass of each precursor i = A,B and R
g
is the gas constant.
We now deﬁne
[S] = precursor A surface concentration (mol/m
2
)
with maximum value [
ˆ
S]. For this mechanism, we compute [
ˆ
S] from the ﬁlm density and measured
growth per cycle under ideal ALD conditions (G
id
pc
). Recall that the [
ˆ
S] value corresponds to maximum
density of sites onto which precursor A can adsorb. Using Al
2
O
3
ALD as an example, we ﬁrst determine
the number density of the AlO
3/2
unit in the amorphous ﬁlm by
n
d
=
ρN
A
M
w
where N
A
is Avogadro’s number, ρ is the ﬁlm density, and M
w
the molecular mass of the ﬁlm minimal
molecular unit (e.g., AlO
3/2
). A ﬁlm monolayer thickness d
z
then can be computed as well as the
fraction F
m
of a monolayer deposited during each ALD cycle as
d
z
= n
−1/3
d
and F
m
=
G
id
pc
d
z
respectively, which leads directly to [
ˆ
S], the quantity that also deﬁnes the maximum number of moles
of each precursor consumed per unit area during each ALD cycle
[
ˆ
S] =
F
m
d
2
z
.
In this manner, we see that F
m
∈ [0, 1] (for “normal” ALD) is the fraction of true reaction sites that can
adsorb the precursor before the surface saturates and that G
id
pc
can be recovered from F
m
d
z
= [
ˆ
S]d
3
z
=
G
id
pc
.
11
Exposure A
With sticking coeﬃcient f
A
∈ [0, 1], a material balance over ∆
t
for the adsorption process for
precursor A gives
[S]
t+∆t
−[S]
t
≈ η
A
o
f
A
_
1 −
[S(t)]
[
ˆ
S]
_
∆
t
where the fractional surface coverage θ for precursor A is deﬁned by
θ(t) =
[S(t)]
[
ˆ
S]
and so (1θ(t))f
A
deﬁnes the probability that a precursor molecule will land on a vacant site at time t
and be adsorbed. As ∆
t
→0, for exposure A we ﬁnd
dθ
dt
=
η
A
o
[
ˆ
S]
f
A
(1 −θ) =
η
A
o
[
ˆ
S]
Γ
A
(θ) (7)
θ(t) = 1 + (θ
o,A
−1) exp
_
−
η
A
o
f
A
[
ˆ
S]
t
_
with θ
o,A
= θ(t = 0). For ideal, fully saturating ALD θ
o,A
= 0. The dosage δ
A
is deﬁned as
δ
A
= P
A
τ
A
=
[
ˆ
S]
_
2πM
A
w
R
g
T
f
A
ln
θ
o,A
−1
θ(τ
A
) −1
with units 1 Langmuir = 10
−6
Torr·s (denoted as L) and where τ
A
corresponds to the length of exposure
A.
Exposure B
Given the stoichiometry of the ALD reaction, we see that the growth surface fully saturated with
precursor A (θ = 1) can be brought back to the state θ = 0 if ν
B
[
ˆ
S] moles/unit area of precursor B
are adsorbed (recall we take ν
A
= 1). While this is not what actually takes place on the surface, we
can still write the absorption site balance under precursor B exposure t ∈ [τ
A
, τ
A
+ τ
B
] starting with
θ
o,B
= θ(τ
A
) as
dθ
dt
= −
η
B
o
ν
B
[
ˆ
S]
f
B
θ = −
η
B
o
ν
B
[
ˆ
S]
Γ
B
(θ) (8)
12
θ(t) = θ
o,B
exp
_
−
η
B
o
f
B
ν
B
[
ˆ
S]
t
_
The corresponding dosage is
δ
B
= P
B
τ
B
=
ν
B
[
ˆ
S]
f
B
_
2πM
B
w
R
g
T ln
θ
o,B
θ(τ
A
+ τ
B
)
with τ
B
corresponds to the length of exposure B.
study Gordon [10] Granneman [11] Rubloﬀ [15] George [14]
ﬁlm HfO
2
HfO
2
HfO
2
Al
2
O
3
precursor A TDMAH TEMAH TEMAH TMA
precursor B water ozone water water
ρ (g/cm
3
) 9.68 9.68 9.68 3.5
GPC (nm/cycle) 0.10 0.10 0.10 0.11
[
ˆ
S] (mol/m
2
) 4.60 ×10
−6
4.60 ×10
−6
4.60 ×10
−6
7.55 ×10
−6
[
ˆ
S] (sites/nm
2
) 2.77 2.77 2.77 4.55
f
i
, i =A,B 1 0.1 0.1 0.1 (to 2 ×10
−5
)
δ
A
ﬂat
(Langmuir) 223 247 240 170 (to 8.5 ×10
5
)
δ
B
ﬂat
(Langmuir) 101 169 101 127 (to 6.4 ×10
5
)
pore R (nm) (93) 100 30 100
L (nm) (8000) 12000 4000 60000
AR 43 60 67 300
P (Torr) (0.00629 ×2) (0.00629 ×2) 0.007 0.01
T (K) 473 500 473 500
τ
A
(s) 0.715 2.03 0.6 210
τ
B
(s) 0.715 2.03 0.6 180
δ
A
pore
(Langmuir) 9000 25500 4200 2.1 ×10
6
δ
B
pore
(Langmuir) 9000 25500 4200 1.8 ×10
6
Table 1: ALD ﬁlm properties, surface reaction conditions, and pore geometries for the four studies to
which our modeling results are compared. Values in parentheses () are estimated from known quantities;
for example the pressure and exposure times corresponding to the Gordon study were computed to
match the known exposure of 9000 L (Langmuir); likewise, the pore dimensions were ﬁtted to the
known AR=43.
Based on θ
o,A
= 0 and taking a representative value of θ(τ
A
) = 0.999 (one open site/1000 at
the end of exposure A) and analogous conditions for exposure B, we compute the absolute minimum
exposure level δ
i
ﬂat
for each precursor; values are listed in Table 1.
13
4.1 Pore surface species balances
A material balance on the pore wall diﬀerential element dA located at axial position z over the time
period ∆
t
during exposure i = A, B gives
ν
i
[
ˆ
S] (θ
t+∆t
−θ
t
) = σ
i
[η
ew
(z) + η
ww
(z) + η
bw
(z) −
w
(z)] ∆
t
= σ
i
[η
ew
(z) + η
ww
(z) + η
bw
(z)] Γ
i
∆
t
where the dimensionless Γ
i
is deﬁned by (7) and (8) and σ
A
= 1, σ
B
= −1. As ∆
t
→0,
ν
i
[
ˆ
S]
∂θ(z, t)
∂t
= σ
i
[η
ew
(z, t) + η
ww
(z, t) + η
bw
(z, t) −
w
(z, t)] .
For the pore bottom
ν
i
[
ˆ
S] (θ
b

t+∆t
−θ
b

t
) = σ
i
[η
eb
+ η
wb
−
b
] ∆
t
= σ
i
[η
eb
+ η
wb
] Γ
i
∆
t
and so we obtain the scalar diﬀerential equation for pore bottom coverage as ∆
t
→0
ν
i
[
ˆ
S]
∂θ
b
(r , t)
∂t
= σ
i
[η
eb
(r , t) + η
wb
(r , t) −
b
(r , t)] .
4.2 Quasi steadystate ﬂuxes
An important assumption in our model is that the emissive and impingement η ﬂuxes are at
equilibrium with respect to the states of the wall and bottom surface coverages, θ and θ
b
, respectively
(we note that this is similar to the assumption used in [18] where the total ﬂux was assumed to be in
equilibrium with respect to the etch surface velocity). This assumption is based on the extremely fast
motions of the precursor species molecules while in transit between pore wall interactions. Therefore,
the overall evolution (slowest timescale for this model) of the surface coverages for highaspectratio
pores is determined by the net ﬂux of precursor species through the pore mouth. Given this equilibrium
(pseudo steadystate) assumption, a material balance on a diﬀerential element on the pore wall and
14
ﬂoor gives (ﬂux in − ﬂux out) = precursor consumed by surface adsorption:
η
ew
(z) + η
ww
(z) + η
bw
(z) −
w
(z) = [η
ew
(z) + η
ww
(z) + η
bw
(z)]Γ
i
(θ(z, t))
η
eb
(r ) + η
wb
(r ) −
b
(r ) = [η
eb
(r ) + η
wb
(r )]Γ
i
(θ
b
(r , t)).
In solving the two equations above given instantaneous values of θ and θ
b
, we ﬁnd we have ﬁve
unknowns: η
ww
(z), η
bw
(z),
w
(z), η
wb
(r ), and
b
(r ). Discretizing the spatially dependent quantities
using n
z
axial and n
r
radial collocation points and corresponding quadrature weights w
z
i
and w
r
j
, we
can write the complete set of equations as follows for i = 1, ... , n
z
and j = 1, ... , n
r
_
1 −Γ
i
(θ(z
i
, t))
_
[η
ew
(z
i
) + η
ww
(z
i
) + η
bw
(z
i
)] −
w
(z
i
) = 0
η
ww
(z
i
) −
nz
k=1
w
(z
k
)Q
ww
(z
i
, z
k
)w
z
k
= 0
η
bw
(z
i
) −
nr
k=1
b
(r
k
)Q
bw
(z
i
, r
k
)w
r
k
= 0
_
1 −Γ
i
(θ
b
(r
j
, t))
_
[η
eb
(r
j
) + η
wb
(r
j
)] −
b
(r
j
) = 0
η
wb
(r
j
) −
nz
k=1
w
(z
k
)Q
wb
(r
j
, z
k
)w
z
k
= 0
with the entrance ﬂuxes explicitly deﬁned as
η
ew
(z
i
) = η
i
o
Q
ew
(z
i
)
η
eb
(r
j
) = η
i
o
Q
eb
(r
j
).
While perhaps more complex in appearance than the single integral equation presented in [13], interesting
features of the ﬂux balance equations are made visible by leaving the ﬂux contributions as separate terms.
15
To illustrate, the set of linear ﬂux balance equations can be written in matrix form
Ax = b with x =
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
η
ww
η
bw
w
η
wb
b
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
,
therefore x ∈ R
3nz +2nr
and the set of linear equations has the structure
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
0 0
0 0 0
0 0 0
0 0 0
0 0 0
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
η
ww
η
bw
w
η
wb
b
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
=
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
0
0
0
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
(9)
where the symbols represent potentially nonzero entries. The oﬀdiagonal block elements of (9) form
the connections between the wall and bottom surface ﬂuxes. We note how the nonhomogeneous terms
are generated only by the inlet ﬂuxes η
ew
and η
eb
, resulting in a trivial solution when η
o
= 0. A trivial
solution also is found for the case f = 1 and θ = θ
b
= 0 during exposure A and θ = θ
b
= 1 during
exposure B; this makes sense because all precursor molecules that enter are consumed by the surface
reactions during the ﬁrst wall/pore bottom collision event.
5 Evolution of θ(z, t) and θ
b
(r , t) during a fullcycle exposure
We now consider a representative simulation in which a 93 nm radius nanopore with aspect ratio 43 is
subject to HfO
2
ALD. The ALD system and pore geometry are chosen to provide a direct comparison to
the classic study of Gordon [10]. A snapshot illustrating the surface coverage dynamics during a TDMAH
exposure is shown in Fig. 4, where the pore membrane top surface (at z = 0) nearly immediately is
saturated with the Hf precursor (compare this ﬁgure with the pore geometry proﬁle shown in Fig. 1,
left). This saturation front moves down the length of the pore during the precursor A exposure period;
our predicted front location closely follows the prediction of the dynamic equivalent to Gordon’s model
16
(taking into account only coating of the pore sidewalls), which gives the saturation front location as
z
G
(t) = 2R
−4 +
_
16 + 6η
i
o
t/(ν
i
[
ˆ
S])
3
. (10)
We compare the inverse of the Gordon model (10) to the maximum possible deposition front position
z
M
(t) computed as
z
M
(t) =
Rη
i
o
t
2ν
i
[
ˆ
S]
(11)
and corresponding to the case that every precursor molecule that enters the pore mouth is adsorbed at
the nearest free surface site.
Figure 4: A snapshot of the pore surface coverage early in exposure A (TDMAH) illustrating the
diminishing surface oxygenligand concentration (in blue) predicted by our ballistic transport/surface
reaction model and compared to the front location predicted by the inverse of the Gordon model z
G
(eqn 10 shown as the green plane) and the maximum surface coverage rate front z
M
(eqn 11, gray
plane) resulting from the adsorption of every precursor molecule entering the pore.
The saturation front dynamics are shown more clearly in Fig. 5. As given in Table 1, the simulation
consists of a 9000 Langmuir exposure to TDMAH followed by a 9000 L water dose. Initially, the pore
inner surface starts with θ(z) = 0 (2.77 sites/nm
2
) available for the Hf precursor adsorption. This is
shown as the solid blue curve in the top plot of Fig. 5. The initial state plus ten evenly spaced θ(z)
proﬁles are shown, illustrating the advancing saturation front during the Hf precursor exposure. Wall
coverages as well as pore bottom proﬁles are shown. In this sequence of plots, we note two diﬀerences
between our simulator predictions and that of the original Gordon paper (the latter’s predictions also
17
0 1000 2000 3000 4000 5000 6000 7000 8000
0
0.5
1
e
exposure A
0 1000 2000 3000 4000 5000 6000 7000 8000
0
0.5
1
e
exposure B
0 1000 2000 3000 4000 5000 6000 7000 8000
0
0.05
0.1
z (nm)
G
P
C
(
n
m
)
0 20 40 60 80
0
0.5
1
e
exposure A
0 20 40 60 80
0
0.5
1
e
exposure B
0 20 40 60 80
0
0.05
0.1
r (nm)
G
P
C
(
n
m
)
Figure 5: Surface coverage dynamics during the TDMAH (exposure A, top) and water (exposure B,
center) half cycles, and resulting GPC as a function of pore depth (bottom) corresponding to the Gordon
[10] study.
are included in Fig. 5): we observe that the pore bottom actually saturates more quickly than the pore
wall adjacent to the pore bottom. This is due to the direct poremouth to bottom ﬂux η
eb
. Second,
the midpoint of our predicted saturation fronts lag slightly behind that of Gordon (eqn 10); this is
attributable to the Langmuirtype adsorption kinetics used in our deposition model.
In this simulation, the Hf precursor nearly saturates the pore; likewise, the water exposure is suﬃcient
to do the same. In Fig. 5 center, the initial condition of the water exposure corresponds to θ = 1 for
most of the pore length. The exposure B (water) front again moved from left to right, more quickly
this time due to the greater ﬂux of water at the pore mouth relative to TDMAH, returning the pore
surface to θ = 0 over its entire length at the end of the full cycle. We note that the time steps used to
illustrate the dynamics of exposure B are the same length as those corresponding to exposure A.
The ultimate ﬁgure of merit in the ALD process is growthpercycle (GPC), and so the ﬁnal step
of the modeling work is to translate the surface reaction rates of the dynamic simulation to GPC. We
note that in cases of undersaturation, GPC will be a function of spatial position with GPC(z) ≤ G
id
pc
.
18
with t = 0, τ
A
as the start and end of exposure A, respectively,
GPC(z) =
_
θ(z, τ
A
) −θ(z, 0)
_
G
id
pc
.
The GPC as a function of pore position for the AB exposure cycle described is shown in the bottom
plot of Fig. 5. Consistent with the predictions of the original Gordon model, the GPC is uniform and at
its nominal value of 0.1 nm/cycle over most of the pore; we note a small degree of precursor starvation
along the pore walls near the bottom and the slightly higher GPC value of the pore bottom itself.
Overall, this simulation corresponds to a nearly ideal TDMAH exposure level and overexposure to the
water precursor.
5.1 Comparisons to other studies
0 2000 4000 6000 8000 10000 12000
0
0.5
1
e
exposure A
0 2000 4000 6000 8000 10000 12000
0
0.5
1
e
exposure B
0 2000 4000 6000 8000 10000 12000
0
0.05
0.1
z (nm)
G
P
C
(
n
m
)
0 50 100
0
0.5
1
e
exposure A
0 50 100
0
0.5
1
e
exposure B
0 50 100
0
0.05
0.1
r (nm)
G
P
C
(
n
m
)
Figure 6: Surface coverage dynamics during the TEMAH (exposure A, top) and ozone (exposure B,
center) half cycles, and resulting GPC as a function of pore depth (bottom) corresponding to the
Granneman [11] study.
Column 2 of Table 1 corresponds to the highaspectratio pore simulations of [11]. In the Granneman
study, the ALD process conditions listed for the AR=60 pore correspond to a 25500 Langmuir exposure
19
of both precursors for this HfO
2
process. As illustrated in Fig. 6, the dynamics of the saturation front
for both halfcycles are qualitatively similar to that described in Fig. 5 except for the more nearly
complete saturation of the full pore length during the Hf precursor exposure in this example. This
leads to practically perfect GPC as a function of pore depth, a result consistent with the conclusions of
Granneman [11]. The nonunity sticking coeﬃcient f
i
= 0.1 used in this simulation further broadens
the front, an observation consistent with [5]. Finally, we note the small deviations from uniformity seen
as a function of radial position r over the pore bottom during both exposures, resulting in a slightly
convex GPC proﬁle on the pore bottom.
0 500 1000 1500 2000 2500 3000 3500 4000
0
0.5
1
e
exposure A
0 500 1000 1500 2000 2500 3000 3500 4000
0
0.5
1
e
exposure B
0 500 1000 1500 2000 2500 3000 3500 4000
0
0.05
0.1
z (nm)
G
P
C
(
n
m
)
0 10 20 30
0
0.5
1
e
exposure A
0 10 20 30
0
0.5
1
e
exposure B
0 10 20 30
0
0.05
0.1
r (nm)
G
P
C
(
n
m
)
Figure 7: Surface coverage dynamics during the TEMAH (exposure A, top) and water (exposure B,
center) half cycles, and resulting GPC as a function of pore depth (bottom) corresponding to the Rubloﬀ
[15] study. Measured GPC data (shown using + markers) are approximated from Fig. 5 of [15].
Column 3 of Table 1 corresponds to the highaspectratio (AR=67) pore simulations of Rubloﬀ [15].
While this is the third HfO
2
ALD case we consider, there are two factors that make this study valuable
for validating our simulation: 1) the system is intentionally underdosed, resulting in partial pore ﬁlling,
and 2) TEM images were used in [15] to directly measure the ALD ﬁlm thickness proﬁle in the region
where precursor depletion presumably begins to take place.
20
Our simulation results are presented in Fig. 7 together with the experimentally measured ﬁlm thick
ness data extracted from [15]. The full pore length of 4 µm is depicted in these plots; we clearly see
that during the TEMAH exposure, the ﬁlm surface saturation front travels only to about onethird of
the full pore length; the subsequent water dose returns the pore to its original fully oxidized state. The
net result is the GPC curve shown in the bottom plot of Fig. 7, where the drop oﬀ in GPC matches
almost perfectly with the measured proﬁle. We note that besides the use of the commonly accepted
value of f
i
= 0.1, there are no adjustable parameters in our overall simulation strategy. The simulation
results verify that it is the subsaturation level of the precursor dosage that leads to precursor starvation
in the depths of the pore.
0 50 100 150 200 250 300 350 400
0
0.2
0.4
0.6
0.8
1
TMA
Water
time (s)
1
ï
e
a
v
g
Figure 8: Integrated normalized surface coverage dynamics during the TMA and water exposures of an
alumina ALD simulation. Data points for surface hydroxyl normalized coverages are approximated from
[14]. Simulator results are shown for f
i
= 0.1 (dashed curve) and f
i
= 2 ×10
−5
(solid curve). We note
that the results for f
i
= 10
−3
are virtually identical to those produced using f
i
= 0.1.
As the ﬁnal example, we examine the potential of our modeling strategy as a tool for probing the
parameters of ALD reaction kinetics. This case corresponds to Column 4 of Table 1 and is the highest
aspect ratio (AR=300) case we consider. In this study of alumina ALD by George and coworkers [14],
the very highaspectratio pores of an alumina membrane provide the necessary substrate surface area
to enable timeresolved FTIR studies of the alumina ALD surface reactions. Because of the high AR
pores, a question we can answer by simulation is whether time delays introduced by precursor transport
down the pores could aﬀect the measurement of intrinsic reaction kinetics.
For this simulation example, we present our results not in terms of surface saturation proﬁles, but as
normalized, poreintegrated values that allow direct comparison to the data of [14]  see Fig. 8. Given
the form of the surface reaction model we use, the normalized average surface hydroxyl coverage is
21
1 −θ
avg
with
θ
avg
=
1
2πRL + πR
2
_
_
L
0
2πRθdz +
_
R
0
2πθ
b
rdr
_
.
Using a sticking coeﬃcient f
i
= 0.1 for both precursors i = A, B (the choice made in the alumina
ALD study of [5]), we see by the dashed curve of Fig. 8 that TMA saturation takes place within 40
s, and water saturation in even less time. We conclude that transport delays may play a small, but
measurable role in terms of interpreting these rate data. Numerical experiments performed in the range
of f
i
∈ [10
−3
, 1] result in nearly identical dynamic results, an interesting fact given the lower range
corresponds to the value used in the Monte Carlo simulations of [6]. Further reduction of the sticking
coeﬃcient to a value of f
i
= 2 × 10
−5
provided the “best” ﬁt to the measured data. The signiﬁcant
diﬀerence between simulator predictions and the published experimental results indicates that further
investigation into the true surface reaction mechanisms and kinetics is required even in the case of
alumina ALD.
6 Conclusions
In this paper we developed a ballistic transport and surface reaction simulator for atomic layer depo
sition processes. Closedform expressions for the intra porefeature precursor transmission probabilities
were developed, compared to previous modeling studies, and coupled to surface reaction models for
binary ALD processes. A contribution of this paper is the detailed discussion presented on the mathe
matical coupling between the transport and surface reaction elements of our approach. A neat division
between the two was uncovered, which will facilitate modifying the surface reaction models or the phys
ical geometry through which precursor transport takes place – identiﬁcation of this model structure
allows changes to be made in one element without disruption to the other.
Our simulation results were compared to four previously published studies of ALD over nanoporous
materials, establishing the validity of our approach using published experimental data. The consistency
of our predictions relative to previous modeling work also was examined; in the one case of alumina
ALD, the signiﬁcant diﬀerences between our model prediction and the experimentally observed precursor
adsorption dynamics pointed to a need for more detailed analysis of the surface reaction kinetics and not
the ballistic transport model. Our goal was to establish a baseline from which subsequent models can be
developed, to explore more intricate deposition surface topographies and more realistic surface reaction
22
mechanisms and kinetic expressions. Furthermore, the basic transport and surface modeling approach
developed in this study and the means by which the spatially varying nature of ﬁlm growthpercycle is
determined will provide the entry point to ALD models that incorporate surface feature evolution, such
as those developed for plasma etch processes [18].
Acknowledgment
The author acknowledges the support of the National Science Foundation through grant CBET
0828410.
References
[1] Adomaitis, R. A., Development of a multiscale model for an atomic layer deposition process, J.
Crystal Growth 312 14491452 (2010), doi: 10.1016/j.jcrysgro.2009.12.041.
[2] Cale, T. S. and G. B. Raupp, Free molecular transport and deposition in cylindrical features, J.
Vac. Soc. Technol. B 8 (4) 649655 (1990).
[3] Cale, T. S, Flux distributions in low pressure deposition and etch models, J. Vac. Soc. Technol. B
9 (5) 25512553 (1991).
[4] Farrington, C. C. Numerical quadrature of discontinuous functions, Proc. 16th Annual ACM Meet
ing 21.40121.404 (1961).
[5] Dendoven, J., D. Deduytsche, J. Musschoot, R. L. Vanmeirhaeghe, and C. Detavernier, Modeling
the conformality of atomic layer deposition: The eﬀect of sticking probability, J. Electrochem. Soc.,
156 (4) 6367 (2009).
[6] Elam, J. W., D. Routkevitch, P. P. Mardilovich, and S. M. George, Conformal coating on untrahigh
aspectratio nanopores of anodic alumina by atomic layer deposition, Chem. Mater. 15 35073517
(2003).
[7] Elliott, S. D. and J. C. Greer, Simulating the atomic layer deposition of alumina from ﬁrst principles,
J. Mater. Chem. 14 32463250 (2004) .
[8] Elliott, S. D., Models for ALD and MOCVD growth of rare earth oxides, Rare Earth Oxide Thin
Films, M. Fanciulli, G. Scarel, Ed., Topics in Appl. Phys. 106 7386 (2007).
23
[9] Gobbert, M. K., V. Prasad, and T. S. Cale, Predictive modeling of atomic layer deposition on the
feature scale, Thin Solid Films 410 129141 (2002).
[10] Gordon, R. G., D. Hausmann, E. Kim, and J. Shepard, A kinetic model for step coverage by ALD
in narrow holes or trenches, Chem. Vap. Deposition 9 (2) 7378 (2003).
[11] Granneman, E., P. Fischer, D. Pierreux, H. Terhorst, and P. Zagwijn, Batch ALD: Characteristics,
comparison with single wafer ALD, and examples, Surface & Coatings Technology 201 88998907
(2007).
[12] Knoops, H. C. M., E. Langereis, M. C. M. van de Sanden, and W. M. M. Kessels, Conformality
of plasmaassisted ALD: Physical processes and modeling, J. Electrochem. Soc. 157 G241G249
(2010).
[13] Kokkoris, G., A. G. Boudouvis, and E. Gogolides, Integrated framework for the ﬂux calculation
of neutral species inside trenches and holes during plasma etching, J. Vac. Sci. Technol. A 24
20082020 (2006).
[14] Ott, A. W., K. C. McCarthy, J. W. Klaus, J. D. Way, and S. M. George, Atomic layered controlled
deposition of Al
2
O
3
ﬁlms using a binary reaction sequence chemistry, Appl. Surf. Sci. 107 128136
(1996).
[15] Perez, I. E. Robertson, P. Banerjee, L. HennLecordier, S. J. Son, S. B. Lee, and G. W. Rubloﬀ,
TEMbased metrology for HfO
2
layers and nanotubes formed in anodic aluminum oxide nanopore
structures, Small 4 12231232 (2008).
[16] Puurunen, R. L., Surface chemistry of atomic later deposition: A case study of the trimethylalu
minum/water system, Appl. Phys. Rev. 97 121301 (2005).
[17] Steckelmacher, W., Knudsen ﬂow 75 years on: the current state of the art for ﬂow of rariﬁed gases
in tubes and systems, Rep. Prog. Phys. 49 10831107 (1986).
[18] Singh, V. K., E. S. G. Shaqfeh, and J. P. McVittie, Simulation of proﬁle evolution in silicon reactive
ion etching with reemission and surface diﬀusion, J. Vac. Sci. Technol. B 10 10911104 (1992).
24
[19] Widjaja, Y. and C. B. Musgrave, Quantum chemical study of the mechanism of aluminum oxide
atomic layer deposition. Appl. Phys. Lett. 80 33043306 (2002).
A Transmission probabilities
As derived in the main body of this paper, the ﬁnal form of the incident ﬂux (molecules/area/time)
along the pore wall resulting from the precursors diﬀusing through the pore entrance is
η
ew
(z) =
_
R
0
_
2π
0
η
o
z(R −r
cos φ
)
π [(r
)
2
+ R
2
−2Rr
cos φ
+ z
2
]
2
dφ
r
dr
= η
o
Q
ew
(z) =
η
o
2R
_
z
2
+ 2R
2
√
z
2
+ 4R
2
−z
_
.
We note that as L →∞, integrating over the entire pore wall surface gives the expected result:
lim
L→∞
_
L
0
_
2π
0
Q
ew
(z) Rdφdz = πR
2
.
We also can observe that
lim
z→0
η
ew
=
1
2
η
o
and lim
z→∞
η
ew
= 0
which also make physical sense.
A.1 Interaction with the pore bottom
A boundary condition is needed at z = L; following Cale [2] and Gordon [10], we consider a ﬂat
pore bottom. We can determine the ﬂux η
eb
directly using φ = 0 without any loss of generality to ﬁnd
n = −z (bottom)
n
= z (inlet)
s = (0 −r
sin φ
)x + (−r + r
cos φ
)y + (L −0)z
s
2
= r
2
+ (r
)
2
−2rr
cos φ
+ L
2
and so
η
eb
(r ) =
_
R
0
_
2π
0
η
o
L
2
π[r
2
+ (r
)
2
−2rr
cos φ
+ L
2
]
2
dφ
r
dr
25
= η
o
Q
eb
(r )
where Q
eb
is
Q
eb
(r ) =
1
2
_
_
1 −
L
2
+ r
2
−R
2
_
(L
2
+ r
2
−R
2
)
2
+ 4L
2
R
2
_
_
with
lim
r →R
Q
eb
(r ) =
1
2
_
1 −
L
√
L
2
+ 4R
2
_
.
An overall material balance can be used to check the computations above:
_
L
0
Q
ew
(z) 2πRdφdz +
_
R
0
Q
eb
(r ) 2πrdr = πR
2
an expression valid for all values of L and R.
A.2 Reemission from pore walls
We now consider the eﬀect precursor reemission from another point z
within the pore has on the
diﬀerential area element located at z. We deﬁne the normal vector of this new element and the vector
connecting it to element dA as
n
= −sin φ
x + cos φ
y + 0z
s = (0 −R sin φ
)x + (−R + R cos φ
)y + (z −z
)z.
Designating η
w
(z
) to be the precursor reemission ﬂux normal to a hemispherical surface at radius s
from the location of dA
in mol/area/time, the ﬂux is (again) governed by the cosine distribution
η
w
(z
) =
w
(z
) cos Ω
2πs
2
c
0
dA
(12)
with dA
= Rdφ
dz
, where c
0
= 1/2 and
w
is the “intrinsic” emission rate (mol/area/time) from any
wall surface point within the pore. The ﬂux η
ww
received at the center point (r = R, φ = 0, z) of
diﬀerential area dA = Rdφdz is
η
ww
(R, φ = 0, z) =
_
L
0
_
2π
0
cos Ωη
w
(z
)
26
=
_
L
0
_
2π
0
w
(z
) cos Ω
cos Ω
πs
2
Rdφ
dz
. (13)
With n = y, n = 1, n
 = 1, and
n · s = R(cos φ
−1)
n
· s = R sin
2
φ
+ R(cos φ
−1) cos φ
= R(1 −cos φ
)
s
2
= R
2
sin
2
φ
+ R
2
(cos φ
−1)
2
+ (z −z
)
2
= 2R
2
(1 −cos φ
) + (z −z
)
2
so
cos Ω = cos Ω
=
R(1 −cos φ
)
[2R
2
(1 −cos φ
) + (z −z
)
2
]
1/2
and
η
ww
(z) =
_
L
0
_
2π
0
w
(z
)R
3
(1 −cos φ
)
2
π[2R
2
(1 −cos φ
) + (z −z
)
2
]
2
dφ
dz
(14)
=
_
L
0
w
(z
)Q
ww
(z, z
)dz
(15)
with
Q
ww
(z, z
) =
_
2π
0
(1 −cos φ
)
2
4πR[(1 −cos φ
) + (z −z
)
2
/(2R
2
)]
2
dφ
=
1
4R
_
_
2 −
3z −z

_
(z −z
)
2
+ 4R
2
+
z −z

3
((z −z
)
2
+ 4R
2
)
3/2
_
_
a result that, again, is consistent with the transmission probabilities listed by Cale [2]. Note that as
L →∞
_
L
0
Q
ww
(L/2, z
) dz
→1 and
_
L
0
Q
ww
(0, z
) dz
=
_
L
0
Q
ww
(L, z
) dz
→
1
2
.
27
A.3 Wallbottom interactions
Finally, we need the ﬂux relationships between the pore bottom and pore wall. For the case of the
walltobottom ﬂux, we consider the bottom diﬀerential element ﬁxed at φ
= 0 and so
n = −z
n
= −sin φ
x + cos φ
y
s = (0 −R sin φ
)x + (−r + R cos φ
)y + (L −z
)z
s
2
= r
2
+ R
2
−2Rr cos φ
+ (L −z
)
2
with
cos Ω =
L −z
_
r
2
+ R
2
−2Rr cos φ
+ (L −z
)
2
cos Ω
=
R −r cos φ
_
r
2
+ R
2
−2Rr cos φ
+ (L −z
)
2
so
η
wb
(r ) =
_
L
0
_
2π
0
w
(z
)(L −z
)(R −r cos φ
)Rdφ
dz
π [r
2
+ R
2
−2Rr cos φ
+ (L −z
)
2
]
2
=
_
L
0
w
(z
)Q
wb
(r , z
)dz
with
Q
wb
(r , z
) =
2R
2
(L −z
)
_
(L −z
)
2
+ R
2
−r
2
¸
_
[(L −z
)
2
+ R
2
+ r
2
]
2
−4R
2
r
2
_
3/2
.
We note that the transmission probability Q
wb
(r , z
) does not have a unique limit where the pore wall
and bottom meet:
lim
r →R,z
=L
Q
wb
=
2R
2
[(L −z
)
2
+ 4R
2
]
3/2
and lim
z
→L
Q
wb
= 0.
If
b
is the intrinsic precursor emission ﬂux at the pore bottom surface, in calculating the bottom
towall ﬂux (from bottom diﬀerential element dA
(r
, φ
) to wall dA(φ, z)), we have a geometry similar
(but opposite in sign with respect to z) to the inletwall case, and so n = y (ﬁxing φ = 0), n
= −z,
28
and
s = (0 −r
sin φ
)x + (−R + r
cos φ
)y + (z −L)z
s
2
= (r
)
2
+ R
2
−2Rr
cos φ
+ (z −L)
2
thus,
η
bw
(z) =
_
R
0
_
2π
0
b
(r
)(L −z)(R −r
cos φ
)dφ
r
dr
π [(r
)
2
+ R
2
−2Rr
cos φ
+ (z −L)
2
]
2
=
_
R
0
b
(r
)Q
bw
(z, r
)r
dr
(16)
where
Q
bw
(z, r
) =
2R(L −z)
_
(L −z)
2
+ R
2
−(r
)
2
¸
_
[(L −z)
2
+ R
2
+ (r
)
2
]
2
−4R
2
(r
)
2
_
3/2
.
In the case where
b
is constant, integrating (16) gives a Q
bw
(z) that is a mirror image of the Q
ew
(z)
plotted in Fig. 2. As with Q
wb
, the limit at the pore wall/bottom junction is not deﬁned:
lim
r
→R,z=L
Q
bw
=
2R
[(L −z)
2
+ 4R
2
]
3/2
and lim
z→L
Q
bw
= 0.
A.4 Fouriersine series  Lobatto quadrature
Each of the subintervals described in the discretization section of the paper are deﬁned by a, b ∈ [0, L]
with b > a and
z
k
= a +
b −a
n
z
−1
k k = 0, 1, ... , n
z
−1.
We expand our unknown function f (z) in the Fouriersine series:
f
nz
(z) =
nz
k=0
c
k
ψ
j
(z) = c
0
_
b −z
b −a
_
+
nz −2
k=1
c
k
sin kπ
_
z −a
b −a
_
+ c
nz −1
_
z −a
b −a
_
. (17)
At the collocation points z, (17) can be written in matrix form as
f = Pc with P
i ,k
= ψ
k
(z
i
).
29
Integrating (17) we ﬁnd our rowarray w of quadrature weights to be
w =
b −a
2
_
1
2 (1 −cos j π)
j π
1
_
P
−1
j = 1, ... , n
z
−2.
We note that interpolation between the diﬀerent quadrature grids is simple to write in terms of matrix
operations. For example, accurate computation of the wall emission contribution to precursor ﬂux out
of the pore at z = L requires interpolating
w
(z) from the coarse to ﬁne grid:
w
(¯z) =
¯
PP
−1
w
(z) with
¯
P
i ,k
= ψ
k
(¯ z
i
).
A.5 Bessel series  Radau quadrature
For this case we consider an interval r ∈ [0, R] with n
r
≥ 2 collocation points deﬁned by
{λ
k
: J
0
(λ
k
) = 0} k = 1, ... , n
r
and r
k
= λ
k
R/λ
nr
. Note that a collocation point at r = 0 is not included because the trial functions
we use are symmetric at this point. We expand our unknown function f (r ) in the FourierBessel series:
f
nr
(r ) =
nr
j =1
c
j
φ
j
(r ) =
nr −1
j =1
c
j
J
0
(λ
j
r /R) + c
nr
_
r
2
j
/R
2
_
. (18)
At each collocation point r
k
, (18) again can be written in matrix form as f = Pc but with P
k,j
= φ
k
(r
j
).
Integrating (18),
_
R
0
f
nr
(r ) rdr =
nr −1
j =1
c
j
λ
j
J
1
(λ
j
)R
2
/λ
2
j
+ c
nr
R
2
4
.
Therefore, our rowarray w
r
of quadrature weights is
w
r
=
_
R
2
λ
j
J
1
(λ
j
)
R
2
4
_
P
−1
j = 1, ... , n
r
−1.
None of the transmission probabilities related to the pore bottom has discontinuous derivatives, so while
we deﬁne a ﬁne grid in r to accurately evaluate the ﬂux integral equations, we do not deﬁne quadrature
grids on subdomains in r ∈ [0, R].