You are on page 1of 16

Chemical Engineering Science 96 (2013) 71–86

Contents lists available at SciVerse ScienceDirect

Chemical Engineering Science


journal homepage: www.elsevier.com/locate/ces

A model-based methodology for the analysis and design


of atomic layer deposition processes—Part III: Constrained
multi-objective optimization
A. Holmqvist a,n, T. Törndahl b, S. Stenström a
a
Department of Chemical Engineering, Lund University, P.O. Box 124, SE-221 00 Lund, Sweden
b
Ångtröm Solar Center, Solid State Electronics, Uppsala University, P.O. Box 534, SE-751 21 Uppsala, Sweden

H I G H L I G H T S

 Novel methodology for the constrained multi-objective optimization of ALD reactor models.
 Analysis of the manufacture of ZnO thin films with constraints on the film thickness uniformity.
 Case study of incommensurable quasi-steady-state reactor throughput and conversion objectives.
 The purge time is an essential parameter in the optimization of throughput in temporal ALD.
 The inherent process robustness is significantly reduced with precursor conversion.

art ic l e i nf o a b s t r a c t

Article history: This paper presents a structured methodology for the constrained multi-objective optimization (MO) of a
Received 20 November 2012 continuous cross-flow atomic layer deposition (ALD) reactor model with temporal precursor pulsing. The
Received in revised form process model has been elaborated and experimentally validated in the first two papers of this series
11 March 2013
(Holmqvist et al., 2012, 2013). A general constrained MO problem (MOP) was formulated to simulta-
Accepted 25 March 2013
neously optimize quasi-steady-state reactor throughput and overall precursor conversion for the
Available online 10 April 2013
controlled deposition of ZnO films from ZnðC2 H5 Þ2 and H2 O, subject to a set of operational constraints.
Keywords: These constraints included lower bounds for the cross-substrate film thickness uniformity and post-
Atomic layer deposition precursor purge duration. The non-dominated Pareto optimal solutions obtained successfully revealed
Optimization
the relation between the incommensurable process objectives and reduced the design space of the ALD
Uncertainty and sensitivity analysis
process into a feasible set of design alternatives. The results presented here show that post-precursor
Numerical analysis
Mathematical modeling purge duration is essential when optimizing throughput in temporally separated ALD processes, and that
Transport processes this is a major drawback when considering operation at atmospheric pressure. Finally, the robustness of
the process along the Pareto optimal front, i.e. the ability of the process to accommodate variations in the
associated set of optimal decision variables (DVs), was assessed by Monte Carlo simulations, in which the
values of the parametric uncertainties were randomly generated from a multivariate normal distribution.
The uncertainty and sensitivity analysis showed that the inherent robustness of the process is
progressively lost with the precursor conversion, and revealed the mechanistic dependence of all DVs
on the proposed optimization specifications.
& 2013 Elsevier Ltd. All rights reserved.

1. Introduction generic ALD process technologies which both rely on decoupling self-
terminating (Puurunen, 2005) gas–surface reactions:
Atomic layer deposition (ALD) is a gas phase deposition technique
that can produce conformal thin films with controlled, uniform (i) Conventional thermal ALD, in which the substrates are sta-
thickness in the nanometer range (George, 2010). There are two tionary and precursors are injected non-overlapping in a time
sequence, separated by intermediate purge steps (Granneman
n
et al., 2007).
Corresponding author. Tel.: +46 46 222 8301; fax: +46 46 222 4526.
(ii) Spatial ALD (S-ALD) (Poodt et al., 2012) with rigid (Fitzpatrick
E-mail addresses: anders.holmqvist@chemeng.lth.se (A. Holmqvist),
tobias.torndahl@angstrom.uu.se (T. Törndahl), et al., 2012; Levy et al., 2009; Nelson et al., 2012) or flexible
stig.stenstrom@chemeng.lth.se (S. Stenström). substrates (roll-to-roll ALD) (Maydannik et al., 2012), in which

0009-2509/$ - see front matter & 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.ces.2013.03.061
72 A. Holmqvist et al. / Chemical Engineering Science 96 (2013) 71–86

continuous precursor flows are separated spatially in different reactor throughput and precursor utilization under the con-
zones and the substrate moves relative to the precursor straints of ensuring spatial control of the cross-substrate film
sources. thickness uniformity and of assigning sufficient post-precursor
purge time.
Industrial semiconductor processing has been a major stimulant (ii) To evaluate the robustness of the set of optimal operating
for the development of the ALD process (Ritala and Niinisto, 2009; points that was determined, i.e. to assess the ability of the
Sneh et al., 2002). Novel applications of ALD are expanding beyond process to accommodate variations in the set of optimal
semiconductor processing in several emerging areas, such as decision variables (DVs) and to rank the variables according
surface passivation layers in c-Si solar cells, buffer layers in to their importance.
CuIn1−x Gax Se2 (CIGS) (Bakke et al., 2011), and diffusion barrier
layers in OLEDs and thin film photovoltaics (Carcia et al., 2009). The reminder of this paper is organized as follows. Section 2
This trend brings with a growing necessity for high-throughput presents the continuous cross-flow ALD reactor system and the
and low cost production techniques (Poodt et al., 2012). mathematical process model. Section 3 describes the formulation
In this study, design parameters for cross-flow, low-volume ALD of the constrained MOP, including its objective functions, algebraic
reactor designs with temporal precursor pulsing were evaluated. The constraints, and DVs. Section 4 is dedicated to the basic compo-
aforementioned reactor designs are of substantial interest for the nents of an uncertainty and sensitivity analysis (UA/SA), while
equipment used to manufacture (Henn-Lecordier et al., 2011) large- Section 5 outlines the modeling and optimization environment.
surface-area substrates (Sundaram et al., 2010). Such substrates are Section 6 presents the results from the multi-objective optimiza-
subject to stringent uniformity, defined by Cleveland et al. (2012) as tion (MO) analysis and the UA/SA of the ALD process model.
the ratio between the standard deviation and the mean value of the Finally, Section 7 contains concluding remarks.
spatially dependent film thickness profile, constraints, while
demands on enhanced throughput (Mousa et al., 2012) and overall
precursor conversion (Lankhorst et al., 2007) require optimal reactor
designs and process operating conditions (Elers et al., 2006). Opti- 2. The continuous cross-flow ALD reactor system
mized reactor design parameters can enable very high throughput,
specifically in the growing number of applications where only very 2.1. Process overview
thin films are needed (Jur and Parsons, 2011). Moreover, optimized
design parameters also allow a significant reduction in precursor The continuous cross-flow ALD reactor system F-120 manufac-
consumption and consequently in the cost-of-ownership, especially tured by ASM Microchemistry Ltd. (Suntola, 1992) was used for the
since it is recognized that many ALD processes under development controlled deposition of thin ZnO films on soda lime glass
for new industrial applications compensate for non-ideal reactor substrates from ZnðC2 H5 Þ2 and H2 O precursors, shown schemati-
designs by significantly over-exposing the substrate to precursors cally in Fig. 1. The reactor setup incorporates an actuator sub-
(Adomaitis, 2011). system that provides flow control (FC) for the carrier gas and each
Wolden et al. (2011) pointed out that the development of precursor inflow, Q_ α and α∈fN2 ; ZnðC2 H5 Þ2 ; H2 Og, and enables
competitive manufacturing techniques requires sophisticated alternate injection with variable dose time, Δt α , of the αth
modeling to understand how to maintain uniformity with respect precursor. Mathematically, the normalized boxcar function,
to both space and time. Furthermore, Henn-Lecordier et al. (2011) Π α ðt; Δt α Þ∈½0; 1, was used to model non-overlapping precursor
and Cleveland et al. (2012) emphasized that the identification of injections in a cyclic time sequence, and is represented as
an ALD process design space (Aelion et al., 1991) defined as the 8
< 1 ϖ α ∑ Δt β ≤t ≤ϖ α ∑ Δt β þ Δt α
multidimensional process design parameter space within which Π α ðt; Δt α Þ ¼ β≠α β≠α ð1Þ
thickness uniformity across the substrate is guaranteed, is non- :
0 otherwise
trivial due to the complex interdependency between the process
design parameters, and requires knowledge of several factors. in which t ¼ ðnΔt −⌊nΔt ⌋ÞΔt is the normalized cycle time,
These include the characteristic time-scales, see e.g. (Adomaitis, nΔt ¼ ðtΔtÞ−1 is the cycle number, Δt ¼ Δt ZnðC2 H5 Þ2 þ Δt N2 þ Δt H2 O þ
2010; Granneman et al., 2007), of reactor-scale mass transport Δt N2 denotes a complete ALD cycle, and ½ϖ ZnðC2 H5 Þ2 ; ϖ H2 O  ¼ ½0; 1.
(Aarik et al., 2006; Jur and Parsons, 2011; Mousa et al., 2012), Furthermore, the continuous inert gas flow, Q_ N2 , transports
and the gas–surface reaction mechanism (Deminsky et al., 2004). traveling waves of adsorptive precursors laterally across the active
However, Londergan et al. (2002) observed that optimal process surface of the substrates, which are mounted on opposite sides of
parameters typically fall outside of the normal operating range the reactor chamber. Temperature control (TC) is achieved through
or control space of an ALD process, in which the impact of an external induction heating setup and ensures that the operating
process-induced variations on film thickness uniformity is more temperature, T, is homogenous, while the pressure, p, at the
pronounced. reactor chamber outlet is controlled by a rotary vane vacuum
The overall objective of this study was to develop a novel, pump (PC). Further details of the reactor system setup are given in,
generally applicable method for optimizing a continuous cross- e.g. Holmqvist et al. (2012), Yousfi et al. (2000), and Baunemann
flow ALD reactor model able to handle multiple objectives and (2006).
constraints. This methodology developed was subsequently
applied to an experimentally validated mechanistic model of the
continuous cross-flow ALD reactor system F-120 manufactured by
ASM Microchemistry Ltd. Suntola (1992) and its development is
described in the first two parts of this paper series (Holmqvist
et al., 2012, 2013). These papers present also a case study of the
deposition of thin ZnO films from ZnðC2 H5 Þ2 and H2 O. The present
paper focuses on two main points:
Fig. 1. Simplified schematic diagram of the continuous flow ALD reactor system
(i) To formulate a constrained multi-objective optimization pro- F-120 manufactured by ASM Microchemistry Ltd. (Suntola, 1992). The vector of
blem (MOP) with the objectives of simultaneously maximizing manipulated variables is: u ¼ ½T; p; Q_ N2 ; Δt N2 ; Q_ ZnðC2 H5 Þ2 ; Δt ZnðC2 H5 Þ2 ; Q_ H2 O ; Δt H2 O † .
A. Holmqvist et al. / Chemical Engineering Science 96 (2013) 71–86 73

where Cieq ðÞ∈Rnk and Ceq ðÞ∈Rnl represent inequality and equality
algebraic constraints, respectively, and the vector of decision
variables, u∈Rnj , is subject to lower and upper bounds that act
as inequality constraints. Incompatibility between the objective
functions and their incommensurability makes it impossible to
Fig. 2. Sketch of the computational domain, Ω, with a partitioned boundary
Γ ¼ Γ in ∪Γ wall ∪Γ sub ∪Γ sym ∪Γ out as outlined in (Holmqvist et al., 2012). The dimen-
find a single ideal solution that is optimal with respect to all
sions are as follows: jΓ sub j ¼ 5:0  10−2 (m), jΓ in j ¼ jΓ out j ¼ 2:0  10−3 (m), and objectives simultaneously. Instead, there is an entire set of solu-
jΓ wall j ¼ 2:5  10−3 (m). tions that are equally optimal, known as Pareto optimal solutions
(i.e. non-dominated solutions), and this set provides flexibility for
2.2. The mathematical process model the decision maker (Miettinen, 1998).

Modeling of an ALD continuous flow reactor system may be 3.1.1. Pareto terminology and the dominance relation
performed at different levels of detail and assumptions, see e.g. Aarik The most common notation of optimality adopted in MO is the
and Siimon (1994), Knoops et al. (2011), Siimon and Aarik (1995, Pareto optimality. The formal definition of the Pareto optimality is:
1997), Yanguas-Gil and Elam (2012), and Ylilammi (1995). The level a decision vector un ∈U is denoted Pareto optimal with respect to
that is chosen depends on the predetermined goal of the modeling, the feasible region
as defined by Hangos and Cameron(2001), and the application of the
model. However, in this study, the experimentally validated mechan- U ¼ fun : Cieq ðx; un ; wÞ ≤0∧Ceq ðx; un ; wÞ ¼ 0g ð4Þ
istic model based on computational fluid dynamics, in which the if there is no another decision vector u∈U for which f ðuÞ ¼
narrow flow channel inside the reactor chamber in the ALD con- ½f 1 ðuÞ; …; f ni ðuÞ† dominates f ðun Þ ¼ ½f 1 ðun Þ; …; f ni ðun Þ† (Miettinen,
tinuous flow reactor system (Fig. 1) constitutes the computational 1998). In this case, a vector f ðuÞ is said to dominate the vector f ðun Þ
domain Ω (Fig. 2), developed in the first two parts of this paper series (denoted f ðuÞ≼f ðun Þ) if the following two conditions are satisfied:
(Holmqvist et al., 2012, 2013) was considered. The process model was
∀i∈f1; …; ni g : f i ðuÞ ≤f i ðun Þ ∧
elaborated and validated experimentally such that it can adequately
predict the substrate, Γ sub , spatially dependent film thickness and ∃ι∈f1; …; ni g : f ι ðuÞ o f ι ðun Þ ð5Þ
provide statistically reliable least-square estimates, β, of the para- n
The concept of dominance, for a given MOP f i ðu Þ, allows to define
meters involved in the heterogeneous gas–surface reaction mechan- the non-dominated, Pareto optimal solution set, P n ∈U (Laumanns
ism that governs the growth of the ZnO film. Thus, no finer modeling et al., 2002)
level of the spatial evolution of the substrate film thickness profile,
P n ¼ fun ∈U : ⇁∃u∈U; f ðuÞ≼f ðun Þg ð6Þ
hs ðζÞ and ζ∈½0; jΓ sub j, than mass and momentum conservation
equations, while incorporating reaction rate equations, was consid- and the decision vector un is termed Pareto optimal if un is non-
ered, which is appropriate for the analysis, design and optimization dominated regarding the entire feasible parameter space U
of ALD reactors (Elliott, 2007). (Eq. (4)). Furthermore, the set of objective vectors f ðun Þ corre-
The model partial differential equations (PDEs) were approxi- sponding to a set of Pareto optimal parameter vectors un ∈P n is
mated using the method of lines (Davis, 1984; Schiesser, 1991) and denoted Pareto optimal front (Veldhuizen and Lamont, 2000), and
the finite element method (Zienkiewicz and Taylor, 2000a,b). For defined as
notational simplicity, the associated system of index-1 differential
PF n ¼ ff ðun Þ : un ∈P n g ð7Þ
algebraic equations (DAEs), FðÞ, in which x, u, and w represent
dependent states, free design variables, and algebraic variables,
respectively, can be written in the general form 3.2. Formulation of objective functions
 
dx
0¼F ; x; u; β; w; t ; xðt 0 Þ ¼ x0 In the present study, the industrial goal of increasing through-
dt
put without compromising the conversion of precursors described
y ¼ gðx; u; β; wÞ
in Section 1 motivated the formulation of the objective functions.
x ¼ ½p; v; ωα ; θκ ; ms † ; ∀α; κ In addition, the developed heterogeneous gas–surface reaction
u ¼ ½T; 〈p〉j ; Q_ ; Δt α † ; ∀α
Γ out α ð2Þ mechanism in the quasi-steady-state growth region (Holmqvist
et al., 2012, 2013) implies that the model response is repre-
where gðÞ is the response function and y denotes the model
sented by limit-cycle solutions and associated with constant
output.
growth per cycle (GPC) (Puurunen, 2003, 2004; Yim et al., 2008).
For this reason, the objective functions and algebraic constraints
3. The multi-objective optimization problem are resolved at the level of a single pulse sequence, Δt. Hence, the
ALD reactor productivity, PjΔt , is defined as the accumulated mass
3.1. Formulation of the constrained multi-objective optimization deposited per unit area of the substrate and cycle time
Z Z
problem 1 ∂ms
PjΔt ¼ ∂Γ ∂t ð8Þ
ΔtjΓ sub j Δt Γ sub ∂t
A MOP involves the simultaneous optimization of i∈f1; …; ni g
where ms is governed by the sum over the sequential hetero-
incommensurable objective functions f ðuÞ ¼ ½f 1 ðuÞ; …; f ni ðuÞ† ∈Rni
geneous gas–surface reaction rates multiplied by the associated
(Deb, 2001; Yee et al., 2003). A general MOP global minimum
difference in molecular weight of the outermost surface species, θκ
problem is formally defined as
(Holmqvist et al., 2012). Moreover, the αth adsorptive precursor
minn −f ðuÞ yield is defined as the relative amount of the inlet precursor dose,
u∈R j
Q_ α Π α ðt; Δt α Þ (cf. Fig. 1), chemisorbed onto the active substrate
subject to Eq: ð2Þ
surface
0≥Cieq ðx; u; wÞ
Z Z Z −1 !
0 ¼ Ceq ðx; u; wÞ jΓ j
Y α jΔt ¼ 1− in N_ α jΓ ∂Γ N_ α jΓ ∂Γ ∂t ð9Þ
umin ≤u ≤umax ð3Þ Δt jΓ out j Γ out Γ in
74 A. Holmqvist et al. / Chemical Engineering Science 96 (2013) 71–86

Table 1 Stokes equations (Holmqvist et al., 2012) in the spatial domain, Ω,


The decision variables, uj and j∈f1; …; 5g, involved in the MOP (Eq. (3)) and their and the second integral of Eq. (10) was derived to assist in the
lower and upper boundaries. The maximum injected αth precursor dose,
physical interpretation of the optimization results. In addition, the
maxðQ_ Π α ðt; Δt α ÞÞj , corresponds to 1:94  102  ΛjΓ j2 (mol), determined with
α Γ in sub
temperature is the most readily accessible attribute to activation
Λ ¼ 1:53  10−5 ðmol m−2 Þ (Holmqvist et al., 2013).
energies of the elementary gas–surface reactions (Lei et al., 2006).
j Decision variables u Units Lower limit umin Upper limit umax There are two main ways of varying the half-cycle average
substrate exposure dose for the αth adsorptive precursor
1 T (K) 3:0  10 2
6:0  102 Z Z
1
2 Q_ (sccm)a 1:0  10 0
2:0  103 〈δα 〉 ¼ p j ∂Γ ∂t ð11Þ
N2
jΓ sub j Δtα Γsub α Γ
3 〈p〉jΓ out (Pa) 5:0  101 1:0  105
4 Q_ ðC2 H5 Þ
Zn 2
(sccm) 0:0  100 2:0  101 either by changing the mass flow of the precursors, Q_ α , thereby
5 Δt ZnðC2 H5 Þ2 (s) 1:0  10−1 1:0  100 changing the partial pressure, or by changing the duration of the
a
pulse, Δt α . Moreover, given a specific set of u ¼ ½T; 〈p〉jΓout ; Q_ N2 , the
Standard cubic centimeters per minute at STP (T ¼273.15 (K) and p ¼1.0
required minimum substrate saturation exposure dose of the αth
(atm)).
adsorptive precursor is governed by
Z Z
1
where N_ α jΓ ¼ −n  ðρωα v−ρDαβ ∇ωα Þ is the total mass flux Λ¼ ∑ ξ r ℓ ∂Γ ∂t
jΓ sub j Δt sat;α Γsub ℓ ¼ 1 α;ℓ
ð12Þ
(calculated using Fick's law of binary diffusion (Bird et al., 1960))
of the αth species normal to the boundary Γ (cf. Fig. 2). Thus, the where the ℓth forward gas–surface reaction rate is proportional to
formulated objective functions can be collectively written as the αth precursor partial pressure in the vicinity of the substrate
f ¼ ½PjΔt ; Y α jΔt † and α∈fZnðC2 H5 Þ2 ; H2 Og. However, in this study, surface, i.e. r ℓ ∝pα jΓ sub (Holmqvist et al., 2012; Masel, 1996).
only the yield of the most expensive ZnðC2 H5 Þ2 precursor was However, productivity enhancement dictates that the overall ALD
considered, reducing the objective space to f ∈R2 . cycle time, Δt, must be substantially reduced, and this leads to

3.3. Definition of decision variables (i) A compromise between Δt α and Δt sat;α .


(ii) The need to minimize post-precursor purge times, Δt N2 .
The set of DVs, u∈Rnj , involved in the optimization procedure
(shown in Fig. 1) may be classified into operating condition Adsorptive precursors are normally alternately injected into the
parameters, u ¼ ½T; 〈p〉jΓ out , Q_ N2 , and ALD pulse sequence para- continuous inert carrier gas flow using gas switching valves,
meters, u ¼ ½Q_ ZnðC2 H5 Þ2 , Δt ZnðC2 H5 Þ2 . The decision vector is subject to where the minimum duration of the pulse is determined by the
the lower, umin , and upper, umax , constraints listed in Table 1, and time required for reliable valve actuation, and is of the order of
this further bounds the feasible decision space u∈U in addition to tenths of a second (Londergan et al., 2002). Moreover, the post-
the algebraic constraints, Cieq and Ceq . Furthermore, the exclusion precursor purge times are adjusted to prevent the co-existence of
of Y H2 O jΔt from the objective vector, f , means that the non- precursors in the reactor chamber, thereby avoiding non-ALD
dominated solutions for u ¼ ½Q_ H2 O ; Δt H2 O  will target max PjΔt , reactions and growth mechanisms (Dasgupta et al., 2012).
whereas Y ZnðC2 H5 Þ2 jΔt is insensitive to perturbations in the DVs.
For this reason, the constant values, ½Q_ H2 O ; Δt H2 O  ¼ ½max Q_ ZnðC2 H5 Þ2 ; 3.4. Formulation of algebraic constraints
min Δt ZnðC2 H5 Þ2 , were assigned to the DVs in the optimization
procedure that governs max PjΔt . The implications and validity of Two algebraic constraints are now introduced to ensure spatial
this postulate will be assessed in Section 6. control of the cross-substrate film thickness uniformity, UFjΔt , and
Optimizing productivity and precursor yield (Eqs. (8) and (9)) to ensure that a sufficient post-precursor purge time, Δt N2 , is
require that an optimal dose of precursors is introduced in allowed. These constraints limit the MOP (Eq. (3)) and define the
minimal time into the reactor chamber that is held at optimal feasible decision space, u∈U, (Eq. (4)).
operating conditions. Aarik et al. (2006), Jur and Parsons (2011),
and Mousa et al. (2012) have all independently shown that the DVs 3.4.1. Cross-substrate film thickness uniformity
of the operating conditions comprise the flow rate and pressure of The design of the continuous cross-flow ALD reactor (Section 2)
the carrier gas, and the process temperature. These parameters leads to the trailing edge of the substrate being exposed to more
together govern the hydrodynamic flow field, and consequently, depleted precursor flow than the leading edge, and the cross-
determine the characteristic time constant for the external mass substrate film deposition rate is consequently inhomogeneous
transport of precursors to the substrate surface and the reactor (Granneman et al., 2007; Cleveland et al., 2012; Henn-Lecordier
residence time. The interdependence between the DVs, deter- et al., 2011). A cross-substrate uniformity metric, UFjΔt , (Cheng and
mined from the inherent formalism of the equation of state Hsiao, 2008) which relates the absolute deviation from the mean
ðρ ¼ pMðRTÞ−1 ; M ¼ ð∑∀α ωα M −1 −1
α Þ Þ, governs the integral mean film height, has been defined
value of the velocity field in Ω (cf. Fig. 2) 8 R −1
Z Z < R
1 1 1 1− Γsub jhs ðζÞ−h s ðζÞj∂ζ Γsub hs ðζÞ∂ζ ∃hs ðζÞ 4 0
〈v〉jΩ ¼ v ∂Ω ¼ ∑ρ Q_ Π α ðt; Δt α Þ ∂Ω ð10Þ UFjΔt ¼ ð13Þ
jΩj Ω jΩj Ω ρAΩ ∀α STP;α α :
1 otherwise
For clarification purposes, the reactor system in Fig. 1 utilizes a where hs ðζÞ is the normalized film height, h s ðζÞ is the mean film
continuous viscous flow of nitrogen carrier gas, Q_ N2 , supplied via a height, and ζ∈½0; jΓ sub j is the substrate boundary, Γ sub , local
mass flow controller, see Holmqvist et al. (2012). Consequently, coordinate variable as outlined previously (Holmqvist et al.,
during the carrier gas purge, Π ZnðC2 H5 Þ2 ¼ Π H2 O ¼ 0 (see Eq. (1)) and 2012). This allows the inequality algebraic constraint to be
the second integral of Eq. (10) is exclusively governed by the expressed as
carrier gas mass flux ðρv≡ρSTP;N2 Q_ N2 ðAΩ Þ−1 Þ. Furthermore, it is
C ieq ¼ UFjΔt;min −UFjΔt ð14Þ
evident that 〈v〉jΩ in Eq. (10) is the average mass velocity of all
the components in the mixture. It is, however, noteworthy that the where UFjΔt;min denotes the minimum acceptable degree of film
velocity field, v, in Eq. (2) is governed by the compressible Navier– thickness uniformity. Moreover, it is noteworthy that C ieq ≤0 in
A. Holmqvist et al. / Chemical Engineering Science 96 (2013) 71–86 75

order for a candidate solution to fulfil the formal feasibility (Lecordier, 2009; Londergan et al., 2002). To evaluate the robust-
requirement, see Eqs. (3) and (4). ness of the optimized continuous cross-flow ALD reactor (Section
2), i.e. to examine how it accommodated variations in optimal DVs,
3.4.2. Post-precursor purge duration model-based uncertainty and sensitivity analysis (UA/SA) (Helton
Precursor physisorption occurs to some extent on internal et al., 2006) was carried out. The main aim was to assess the
surfaces throughout the entire reactor system (cf. Fig. 1), main- impact of variations in key parameters in the Pareto optimal
tained at or near the substrate temperature during the exposure. solution set, P n (Eq. (6)), on product quality, measured by the film
Desorption will subsequently occur during the sequential precur- thickness uniformity (Eq. (13)), and on process performance,
sor exposure period if the reactor has not been subjected to measured by productivity and precursor yields (Eqs. (8) and (9)),
sufficient purging by an inert carrier gas. In order to achieve along the Pareto optimal front, PF n (Eq. (7)). Several approaches
sufficient removal of precursors, the characteristic time constant to UA/SA have been developed, however, an approach based on
for the temperature-dependent desorption process (Groner et al., Monte Carlo methods was used in this study.
2004), the reactor system mean residence time (Elers et al., 2006),
and the carrier gas entrainment (Aarik et al., 2006), must be 4.1. Sampling-based methods
assessed to quantify the background precursor level. Jur and
Parsons (2011) and Mousa et al. (2012) suggested that the lower The initial component that underlies the implementation of
bound of the purge time, Δt N2 , needed in the ALD sequence to sampling-based methods involves assigning distributions that
ensure negligible precursor interaction is determined from the characterize the uncertainty in each of the elements of u. When
maximum values of: no information on the variability of an input parameter is avail-
able, a normal distribution, uj ∼N ðμuj ; s2uj Þ, is generally a good
approximation (according to the central limit theorem
(i) The average gas residence time Vardeman, 1994). The subsequent component comprises the gen-
eration of a sample. Latin hypercube sampling (LHS) is suitable for
jΓ sub j þ 2jΓ wall j
τ¼ ð15Þ use with computationally demanding models because its stratifi-
〈v〉jΩ
cation properties ensure that the whole variable range is uni-
for removing species from the reactor by the bulk purge flow, formly sampled, even with a relatively small sample size (Helton
and ultimately determined from the interdependence of Q_ N2 , p and Davis, 2003). The final component explores and illustrates the
and T through Eq. (10). mapping (Helton and Davis, 2002), during which an approxima-
(ii) The mixing time tion of an empirical probability density function (pdf) of the model
Δ2 output provides the most complete representation of the uncer-
τDαβ ¼ ð16aÞ tainty in the elements of y that is derived from individual
Dαβ
distributions in u. The pdf, ϕðXÞ, of the jdimensional multivariate
!1=2 normal distribution of the vector X, collecting elements of u and y,
ηz1
Δ ¼ 4:64 ð16bÞ is given by
ρvz1 jΓsym  j=2  
1 1 1 † −1
which is required for physisorbed precursors and vapor ϕðXÞ ¼ exp − ½X−μX  ΣðXÞ ½X−μ X  ð18Þ
2π ðdet ΣðXÞÞ1=2 2
product species that are generated to traverse the laminar
boundary layer thickness, Δ, over a flat surface into the bulk which is parametrized with a mean vector, μX , and a ðj  jÞ
flow through diffusion (Bird et al., 1960). covariance matrix
0 1
s21 ϱ12 s1 s2 … ϱ1j s1 sj
Hence, the lower boundary of the N2 purge period, Δt N2 , is B C
B ϱ12 s1 s2 s22 … ϱ2j s2 sj C
imposed by the equality constraint. B C
ΣðXÞ ¼ B C ð19Þ
B ⋮ ⋮ ⋱ ⋮ C
C eq ¼ Δt N2 −maxðNΩ τ; τDαβ Þ ð17Þ @ A
ϱ1j s1 sj ϱ2j s2 sj … sj 2

in which N Ω denotes the number of reactor chamber volumes to


be purged. However, given the narrow dimensions of the flow Determining the expectation values and correlation coefficients
channel (Fig. 2) and the DV boundaries that are listed in Table 1, associated with Eqs. (18) and (19) lays the foundation for ranking
τDαβ ⪡τ and, consequently, the purge period, Δt N2 , is exclusively the contributions from individual elements of u with respect to
determined by τ. It is noteworthy that the continuum gas flow the uncertainty in the elements of y.
calculations (Eqs. (15) and (16)) do not consider specific time
constants for the species desorption kinetics. This would ulti-
mately require a reaction mechanism for the desorption of 5. Modeling and optimization environment
physisorbed precursors, which is not considered in the ALD s
process model (Holmqvist et al., 2012, 2013). COMSOL Multiphysics (COMSOL AB, 2008), a finite element
analysis, solver and simulation software package for encoding of
complex and coupled physical systems aimed at simulation, is
4. Uncertainty and sensitivity analysis used for reactor modeling (see Section 2.2). It also offers an
extensive interface to MATLABs (The MathWorks, Inc., 2010a),
The stability and repeatability of the optimized ALD process and hence, is a suitable platform for optimization, to be used in
s
will depend on the impact of variations of the key operating conjunction with COMSOL Multiphysics .
condition parameters on the resulting film quality and process
performance. The industrial goal of increasing throughput and the 5.1. Multi-objective evolutionary algorithms
conversion of precursors dictate the use of very short ALD cycle
times, Δt. This regime is typically outside of a conventional ALD From the class of multi-objective evolutionary algorithms
process design space, in which the surface reactions approach (MOEAs), two state-of-the-art algorithms: NSGA-II (Deb et al.,
saturation as the substrate becomes over-exposed to precursors 2002) and SPEA2 (Zitzler et al., 2001), are widely used, since they
76 A. Holmqvist et al. / Chemical Engineering Science 96 (2013) 71–86

are based on principles of dominance and Pareto optimality


(defined in Section 3.1.1) and they possess the desired convergence
and diversity-preservation properties (Laumanns et al., 2002). The
model-based methodology proposed in this study is based on the
following procedure: the parallelized evolutionary algorithm dif-
ferential evolution (DE) (Price, 1999; Price et al., 2005; Storn and
Price, 1997), following the master–slave parallel and distributed
programming paradigm (Coello et al., 2007) as outlined previously
(Holmqvist et al., 2013), was extended for handling MOPs by
adopting the external elitist archive from the SPEA2 algorithm and
the constraint-handling approach from the NSGA-II algorithm. The
algorithm thus implemented is similar to the existing multi-
objective self-adaptive differential evolution (MOSADE) algorithm
proposed by Wang et al. (2010) in that it adopts the elitist
reservation strategy and the preservation of diversity based on
crowding degree (in the objective space Rni ). However, the self-
adaptation strategy for the DE control parameter derived by Qin
and Suganthan (2005) and Qin et al. (2009) was used.

5.1.1. Assessment of Pareto optimal solution sets


Furthermore, the deterministic weighted-sum-method (Kim
and de Weck, 2005) was used in order to assess the reliability of
the output from the implemented algorithm and the convergence Fig. 3. Projection of the feasible objective space f ¼ ½PjΔt ; Y ZnðC2 H5 Þ2 jΔt † ∈R2 for the
to the true Pareto optimal front, PF n . This method aggregates all constrained MOP (Eq. (3)) subject to the set of DVs and associated boundaries listed
objectives, f ðuÞ∈Rni , into one multi-attribute utility function in Table 1. The Pareto optimal fronts, PF n , are shown for: ð−○−Þ UFjΔt;min ¼ 0:90;
ð−▿−Þ UFjΔt;min ¼ 0:95; ð−◊−Þ UFjΔt;min ¼ 0:99. Pareto optimal solutions generated by
(MAUF), Uðu; λÞ, by a linear combination via the normalized weight
the resolution of Uðun ; λÞ and associated weight vector λ1 ¼ ½0:0; 0:1; …; 1:0† are
vector, λ, as follows: shown for: ðÞ UFjΔt;min ¼ 0:90; ð▾Þ UFjΔt;min ¼ 0:95; ð♦Þ UFjΔt;min ¼ 0:99. ðÞ Domi-
nated solutions for UF min jΔt ¼ 0:90; ð◀Þ Ex situ XRR film thickness measurement
Uðu; λÞ ¼ ∑ λi f i ðuÞ ð20Þ data presented previously (Holmqvist et al., 2012).
i¼1

The convex MAUF was optimized by the indirect single-objective


optimization routine fmincon, in the Optimization Toolbox™ for algebraic constraints are interdependent. Hence, it is not possible
MATLABs (The MathWorks, Inc., 2010b). It is worth noting that to avoid sub-optimal solutions when using the sequential one-
Uðu; λÞ is the convex function of normalized objectives, f i variable-at-a-time (OVAT) approach. In contrast, the extended
 ϑi multivariate MO method presented in Section 3 makes no
max f i ðuÞ−f i ðuÞ
f i ðuÞ ¼ ð21Þ assumptions about the structure of ΣðXÞ, and this method is
max f i ðuÞ−min f i ðuÞ
therefore more appropriate for determining the optimal decision
in which ϑi are the relative tolerances (Benyahia et al., 2011) and space, un ∈P n .
the weights λi are chosen such that λi ≥0 and ∑i ¼ 1 λi ¼ 1. For
further details and geometric interpretation in the objective 6.1. Multi-objective optimization analysis
function space of the weighted-sum-method see e.g. Logist et al.
(2009). The proposed model-based methodology for the multivariate
and constrained MOP gives adequate importance to the incom-
mensurable objective functions (Eqs. (8) and (9)) under the
6. Results and discussion constraints of film thickness uniformity (Eq. (14)), purge duration
(Eq. (17)), and set of DV boundaries (Table 1). Fig. 3 presents the
Previous work on optimization and design strategies for ALD approximation of the Pareto optimal front, PF n , for UFjΔt;min ∈
processes has mainly focused on throughput enhancement. This is ½0:90; 1:0½ and N Ω ¼ 2:0 obtained with the implemented MOEA,
true of both theoretical work (Lim et al., 2000, 2001; Park et al., subject to Eq. (3). The figure presents also the solutions generated
2000) and experimental studies (Aarik et al., 2006; Jur and by the fmincon algorithm (The MathWorks, Inc., 2010b) subject to
Parsons, 2011; Kuse et al., 2003; Lei et al., 2006; Mousa et al., Eq. (20) and the associated weight vector λ1 ∈½0:0; 1:0 and relative
2012). The strategies developed in this way are primarily based on tolerance vector ϑ ¼ ½5:0; 5:0  10−2 . It is noteworthy that the
optimizing the average precursor substrate exposure dose, 〈δα 〉 solutions Uðun ; λÞ are superjacent to the solutions f ðun Þ, validating
(Eq. (11)). Furthermore, the comprehensive univariate sensitivity that the optimization has converged to the true PF n . Furthermore,
analysis of u on mass gain per cycle (MGPC) (Burton et al., 2009), the objective function values, f ¼ ½PjΔt ; Y ZnðC2 H5 Þ2 jΔt , associated
〈ms 〉, and UFjΔt , presented in Holmqvist et al. (2012) extends the with the experimental X-ray reflectivity (XRR) ZnO film thickness
framework for throughput enhancement to include the impact measurement data presented previously (Holmqvist et al., 2012)
of inhomogeneous precursor flux profiles across the substrate clearly show the prospects of optimizing temporally separated
during the exposure period (Gobbert et al., 2002). This exten- ALD processes. It is, however, noteworthy that the experimental
sion complements experimental studies (Cleveland et al., 2012; operating conditions governed the number of reactor chamber
Henn-Lecordier et al., 2011) that assessed the uniformity of film volumes purged in the range N Ω;XRR ∈3:0; 4:0½102 , and thus,
thickness in cross-flow ALD reactor designs. explains the large difference between the experimental deter-
However, the univariate method makes a fundamental assump- mined and the model generated PjΔt .
tion about the structure of the covariance matrix, ΣðXÞ (Eq. (19)), In addition, Fig. 3 shows that the discrepancy between the Pareto
with ϱjj ≡0 and ∀j. This assumption is not adequate in a situation in optimal fronts for UFjΔt;min ∈½0:90; 1:0½ is lower for Y ZnðC2 H5 Þ2 jΔt o
which DVs and relations between the DVs and the objectives and 0:40. As emphasized in Section 3.4.1, the non-uniform film height,
A. Holmqvist et al. / Chemical Engineering Science 96 (2013) 71–86 77

Fig. 4. Film thickness profile, hs ðζÞ, as a function of the substrate boundary, Γ sub , local coordinate variable, ζ∈½0; jΓ sub j, for unλ associated with λ1 ¼ ½0:1; 0:2; 0:4; 0:6† and
−1
UFjΔt;min ∈½0:90; 1:0½. hs ðζÞ (Å cycle ) associated with: (–) UFjΔt;min ¼ 0:90; ð−  −Þ UFjΔt;min ¼ 0:95; (– –) UFjΔt;min ¼ 0:99.

hs ðζÞ, in the range ζ∈½0; jΓ sub j originates from the decrease in gas– increment trajectory, 〈ms 〉 ¼ 〈ms 〉ðΛM s Þ−1 , for unλ and λ1 ¼ 0:9
surface reaction driving force, hence r ℓ ∝pα jΓsub (Holmqvist et al., (Fig. 5a) exceeds 90% of the theoretical maximum deposited ZnO
2012), which in turn results from depletion of the precursor pulse film mass at t ¼ Δt, whereas the associated mean gas–phase molar
along the direction of the flow. This phenomenon is evident when compositions, 〈ψ α 〉, at Γ in and Γ out correspond to low ZnðC2 H5 Þ2
analyzing the spatial resolution of hs ðζÞ for the set of optimal yield, Y ZnðC2 H5 Þ2 jΔt ¼ 14:8%. In contrast, the model response asso-
operating points ciated with unλ and λ1 ¼ 0:1 (Fig. 5d) corresponds to a high
ZnðC2 H5 Þ2 yield, Y ZnðC2 H5 Þ2 jΔt ¼ 96:2%, whereas 〈ms 〉 o 0:70 at
unλ ¼ fun : ⇁∃u∈Rnj ; Uðu; λÞ≥Uðun ; λÞ∧f ðun Þ∈PF n g ð22Þ t ¼ Δt. Hence, λ1 ¼ ½0:0; 1:0 comprises the anchor points of the
solutions of the single-objective optimization problems for each
and associated weight vector, λ1 ¼ ½0:1; 0:2; 0:4; 0:6, depicted in criterion.
Fig. 4, where higher values of Y ZnðC2 H5 Þ2 jΔt lead to lower film thickness
uniformity (Eq. (13)). Consequently, in order to fulfill more stringent
constraints on the uniformity of film thickness (corresponding to 6.1.1. Assessment of the optimal operating condition parameters
UFjΔt;min -1:0), sufficient driving forces in the forward reaction rates, The projection of the feasible objective function space and
r ℓ , must be preserved as ζ-jΓ sub j, which in turn requires higher the feasible inequality algebraic constraint space depicted in Fig. 6
half-cycle average substrate exposure doses, 〈δZnðC2 H5 Þ2 〉 (Eq. (11)), provide rigorous information about the Pareto optimal solution
at the expense of lower Y ZnðC2 H5 Þ2 jΔt . This is apparent when analyzing set, P n ∈U, which comprises the ALD process design space. More-
Fig. 3, where the optimal Pareto front, PF n , in the range over, the distinctive scatter of feasible dominated solutions in
Y ZnðC2 H5 Þ2 jΔt ∈0:0; 0:4½ is forced towards the origin as UFjΔt;min -1:0. Fig. 6 reveals the corresponding dependence in the objective
Fig. 5 shows the dynamic model response resolved at the function space f ∈R2 for UFjΔt;min ¼ 0:90 depicted in Fig. 3. For
level of a single ALD pulse sequence for the set of optimal clarification purposes, the feasible dominated solutions, u∈U, in
operating points, unλ , in order to demonstrate the implication Fig. 6 are distributed between the DV lower and upper boundaries
of a single-objective optimization of the incommensurable objec- that are listed in Table 1, and reveal the relations between the DVs
tives, f ¼ ½PjΔt ; Y ZnðC2 H5 Þ2 jΔt . The normalized integral mean mass and the objective functions, PjΔt and Y ZnðC2 H5 Þ2 jΔt , and the algebraic
78 A. Holmqvist et al. / Chemical Engineering Science 96 (2013) 71–86

Fig. 5. Dynamic process analysis resolved at the level of one ALD pulse sequence, Δt, for UFjΔt;min ¼ 0:90 and unλ associated with λ1 ¼ ½0:1; 0:4; 0:6; 0:9† . (–) 〈ms 〉 ¼ 〈ms 〉ðΛM s Þ−1
(a.u.) at Γ sub . Integral mean gas–phase molar composition 〈ψ α 〉jΓ (a.u.) of the αth precursor ðα∈fZnðC2 H5 Þ2 ; H2 OgÞ at Γ in and Γ out : ð−  −Þ 〈ψ ZnðC2 H5 Þ2 〉jΓ in ; ð○−Þ 〈ψ ZnðC2 H5 Þ2 〉jΓ out ;
(- -) 〈ψ H2 O 〉jΓ in ; ð−▿−Þ 〈ψ H2 O 〉jΓ out .

constraint, UFjΔt . Furthermore, the superposed non-dominated The projection of the non-dominated solutions for Q_ N2 shown in
solution set, P n ∈U, shows the subset of feasible solutions that Fig. 6 exhibits weak convex behavior with respect to PjΔt . Accordingly,
−1
governs the Pareto optimal front, PF n , depicted in Fig. 3. Thus, the a reduction in Q_ N2 increases the inlet mass fraction ðωα jΓin ∝Q_ N2 Þ of
presented multivariate MO method enables simultaneous investi- the αth precursor, and promotes the driving force in the forward
gation of the interdependence of the DVs, and the dependence elementary gas–surface reactions (r ℓ ∝pα (Holmqvist et al., 2012)) at
−1
of the optimal DVs on the proposed objectives and algebraic the expense of an increase in the duration, Δt N ∝Q_ , of the post- 2 N2
constraints.
precursor purge. However, since Y ZnðC2 H5 Þ2 jΔt (Eq. (9)) is time-invar-
The dependence of the feasible dominated solutions for the
temperature, T, (Fig. 6) exhibits convex behavior with respect to iant, maxY ZnðC H Þ j is achieved at the lower boundary of Q_ , as
2 5 2 Δt N2

both PjΔt and Y ZnðC2 H5 Þ2 jΔt . The model response in the range expected. Likewise, the dependence of the non-dominated solutions
T∈½3:0; 4:5½102 (K) is attributed to the activation energy in the for 〈p〉jΓout on Y ZnðC2 H5 Þ2 jΔt is dictated to a high degree by the
Arrhenius equation (Schwaab and Pinto, 2007, 2008; Schwaab aforementioned parametric dependence, and corresponds to an
et al., 2008) of the elementary gas–surface reaction mechanism as optimal gas–surface reaction rate ðr ℓ ∝pα Þ, in which the molar fraction,
defined in Holmqvist et al. (2012). The model response in the ψ , of the αth precursor is ultimately governed by Q_ .
α N2
range T∈4:5; 6:0  102 (K), however, originates from extensive The feasible dominated solutions of 〈p〉jΓout also depend on PjΔt
dehydroxylation, which overrides the effect of the increase in in a convex manner (Fig. 6). The model response in P Δt for 〈p〉jΓout is
hydroxyl groups during the H2 O exposure sequence (Deminsky ultimately determined by the incommensurability between the
et al., 2004; Matero et al., 2000). Additionally, Eqs. (10) and (15)– forward elementary gas–surface reaction rates, r ℓ ∝p, and the post-
(17) make it clear that Δt N2 ∝pðT Q_ N2 Þ−1 , which implies that the precursor pulse duration, Δt N2 ∝p. This is evident from analyzing
optimal T (which corresponds to max PjΔt ) is a compromise the associated post-precursor purge period, which arises from
between a high gas–surface reaction rate and a low Δt N2 . On the the parametric interdependence of u ¼ ½T; 〈p〉jΓout ; Q_ N2  through
other hand, Y ZnðC2 H5 Þ2 jΔt is time-invariant (Eq. (9)), and is not Eqs. (10) and (15)–(17) for the set of non-dominated solutions.
affected by the parametric dependence of Δt N2 . Hence, as 〈p〉jΓout approaches atmospheric pressure, the overall
A. Holmqvist et al. / Chemical Engineering Science 96 (2013) 71–86 79

Fig. 6. Projection of the feasible objective function space and the feasible inequality constraint space for UFjΔt;min ¼ 0:90 and the set of DVs u ¼ ½T; 〈p〉jΓ out ; Q_ N2 , and the
associated post-precursor purge time, Δt N2 (Eqs. (10) and (15)–(17)). ð○Þ Feasible Pareto optimal solution set, P n ∈U; ðÞ Feasible dominated solutions, u∈U; ð◀Þ Ex situ XRR film
thickness measurement data presented previously (Holmqvist et al., 2012).

productivity (Eq. (8)) is favored more by a reduction in Δt N2 than it solutions accumulate near the lower boundary (Table 1). This
is by promoting the forward gas–surface reaction rates. The impact shows that a reduction in Δt N2 favors the overall productivity
of 〈p〉jΓ out on PjΔt for N Ω ∈½2:0; 1:0  102  (Eq. (17)) will be discussed (Eq. (8)) more than it promotes the forward gas–surface reaction
in detail in Section 6.1.2. rates. It is, however, noteworthy that the severe dependence of
In conclusion, the projection of the non-dominated solutions Δt N2 on P Δt falls as N Ω -1:0. This implies that the duration of the
depicted in Fig. 6 demonstrates that it is impossible to optimize post-precursor purge is of key importance when optimizing the
the set of DVs u ¼ ½T; 〈p〉jΓ out ; Q_ N2  using an OVAT approach without throughput of cross-flow, low-volume ALD reactor designs with
obtaining solutions that are essentially of a local optimization temporal precursor pulsing, and that it presents a major obstacle
nature. This is a consequence of the severe parametric interde- when considering the optimal operation of atmospheric pressure
pendence, ultimately determined from the inherent mathematical ALD designs.
structure of Eq. (10).

6.1.3. Assessment of the optimal pulse sequence parameters


6.1.2. Assessment of the constraint on post-precursor purge duration
Fig. 8 shows how the feasible non-dominated solutions,
Fig. 7a shows Pareto optimal fronts, PF n , for N Ω ∈½2:0; 1:0
2 P n ¼ ½Q_ ZnðC2 H5 Þ2 , Δt ZnðC2 H5 Þ2 ∈U, depend on f ¼ ½PjΔt ; Y ZnðC2 H5 Þ2 jΔt . It
10 , and demonstrates the impact of the predefined number of
is noteworthy that the ultimate model response is governed by the
reactor chamber volumes purged, NΩ in Eq. (17), during the post-
product of these two DVs, which represent the precursor dose
precursor purge period, Δt N2 , on the feasible objective space,
injected, ðQ_ α Π α ðt; Δt α ÞÞjΓin . Consequently, the MOEA is incapable of
f ¼ ½PjΔt ; Y ZnðC2 H5 Þ2 jΔt . It is apparent that max Y ZnðC2 H5 Þ2 jΔt ¼ 1:0 is
distinguishing between values of Δt ZnðC2 H5 Þ2 ≠0 when Q_ ZnðC2 H5 Þ2 -0,
achieved for the entire range of N Ω , since this optimization
due to the way in which the parameters depend on each other.
criterion is time-invariant, whereas the response in PjΔt is non-
The projection of the dominated solutions for Δt ZnðC2 H5 Þ2 shown
linear and max PjΔt is reduced by a factor of 5.5.
in Fig. 8, behave in a non-convex manner with respect to both
In addition, the behavior of 〈p〉jΓout with respect to PjΔt (Fig. 7b)
objectives, and the non-dominated solutions accumulate at the
is non-convex for N Ω ¼ 1:0  102 , and the non-dominated lower boundary, min Δt ZnðC2 H5 Þ2 . Furthermore, the non-dominated
80 A. Holmqvist et al. / Chemical Engineering Science 96 (2013) 71–86

it was postulate in Section 3.3 that ½Q_ H2 O , Δt H2 O  ¼ ½max Q_ ZnðC2 H5 Þ2 ,


min Δt ZnðC2 H5 Þ2  governs max PjΔt , and Fig. 8 shows that this is valid
for the ZnðC2 H5 Þ2 precursor.

6.2. Uncertainty and sensitivity analysis

The non-dominated solutions, P n ∈U, shown in Figs. 6 and 8


provide flexibility for the decision maker, but information from the
UA/SA analysis about the robustness along the Pareto optimal
front, PF n , (Fig. 3) is also needed in order to determine the
optimal control space in which the process accommodates input
variability in the optimal set of DVs without volatility in the
assigned constraints (Eq. (4)). It was shown in Section 6.1 that the
inequality constraint of the uniformity of cross-substrate film
thickness (Eq. (14)) dictates to a high degree the extension of
the Pareto optimal front, and this is the reason that the post-
optimization analysis is needed. Consequently, several of the
solutions along the Pareto optimal front possess the minimum
acceptable degree of film thickness uniformity ðUFjΔt ≡UFjΔt;min Þ,
and the existence of these solutions implies that the process is on
the verge of failure.

6.2.1. Uncertainty analysis


In order to evaluate the robustness of the set of non-dominated
solutions, the associated optimal decision space, unλ (Eq. (22)), was
subject to the assigned normal distribution, u^ λ ∼N ðunλ ; s2un Þ, with
λ
lower and upper control limits with a percentage deviation of
sunλ ¼ ½1:0; 5:0; 10:0ð%Þ, in all the DVs listed in Table 1 and with a
limit of 95% of the values drawn from the distribution within two
standard deviations, sunλ , away from the mean, unλ . For each case
λ1 ¼ ½0:1; 0:2; 0:4; 0:5; 0:6; 0:8; 1:0, ju^ λ j ¼ 2:0  103 process scenar-
ios were selected using lhsnorm in the Statistics Toolbox™ for
MATLABs (The MathWorks, Inc., 2010c), and assigned a duration
n
of post-precursor purge, Δt N2 , of u^ λ as specified by Eq. (17). Fig. 9
shows the resulting model response in f ¼ ½PjΔt , Y ZnðC2 H5 Þ2 jΔt  for
½UFjΔt;min ; N Ω  ¼ ½0:95; 2, while Table 2 lists the fraction of the
sample, ju^ λ ∉Ujðju^ λ jÞ−1 that violates the algebraic inequality con-
straint (Eq. (14)) for each case and degree of DV perturbation. In
addition, Fig. 9 shows the projection along the principal axes of
the concentric hyperellipsoid contour with equal probability ð1−αÞ
centered at μ^ X

^
fX : ½X−μ^ X † ΣðXÞ −1
½X−μ^ X  ¼ χ 2ðφ;αÞ g ð23Þ

with the estimated covariance matrix, ΣðXÞ ^ (Eq. (19)), using


gmdistribution in the Statistics Toolbox™ for MATLABs . By this
Fig. 7. Feasible objective space f ¼ ½PjΔt ; Y ZnðC2 H5 Þ2 jΔt † ∈R2 for the constrained MOP
means, it was possible to quantify the magnitude and direction
(Eq. (3)) subject to u and the boundaries listed in Table 1. (a) Pareto optimal front,
PF n , for N Ω ∈½2:0; 1:0  102 . (b) Normalized productivity, P jΔt , as a function of the
of displacements associated with the model response, X ¼ f ðu^ λ Þ.
non-dominated solutions for log10 〈p〉jΓ out . ð○Þ N Ω ¼ 2:0  100 ; ð▿Þ N Ω ¼ 1:0  101 ; It is apparent from Fig. 9 that the eigenvector of the estimated
ð◊Þ N Ω ¼ 2:5  101 ; ð△Þ N Ω ¼ 1:0  102 ; ð◀Þ Ex situ XRR film thickness measurement covariance matrix, Σðf ^ ðu^ λ ÞÞ, approximately intersects the origin
data presented previously (Holmqvist et al., 2012). for each case λ1 . This means that the process uncertainties,
u^ λ ∼N ðunλ ; s2un Þ, display a loss in f in the direction towards the
λ
origin. Accordingly, the difference between the estimated μ^ f ðu^ λ Þ
n
of Eq. (18) and f ðu^ λ Þ has a significant trend along the PF n , where
solutions associated with Q_ ZnðC2 H5 Þ2 are distributed in the range PjΔt is most uncertain as λ1 -1:0 and Y ZnðC2 H5 Þ2 jΔt as λ1 -0:0
½min Q_ ZnðC2 H5 Þ2 ; max Q_ ZnðC2 H5 Þ2 . Accordingly, the optimal injected (cf. Table 2).
precursor dose with respect to PjΔt is governed by max Q_ α min Δt α , Fig. 10 shows in a similar manner the simulated model
whereas min Q_ α min Δt α is optimal with respect to Y α jΔt and response distribution in UFjΔt for each case λ1 and sunλ ¼ 10:0ð%Þ.
α ¼ ZnðC2 H5 Þ2 . This implies that minΔt ZnðC2 H5 Þ2 ≡Δt sat;ZnðC2 H5 Þ2 as It is evident that the robustness of the process increases as the
maxQ_ ZnðC2 H5 Þ2 gives the saturation half-cycle average substrate parameter set moves towards the operating points as λ1 -1:0. This
exposure dose, 〈δZnðC2 H5 Þ2 〉 (Eq. (11)), which satisfies Eq. (12). Once observation is consistent with the findings presented in Section
it has been guaranteed that Eq. (12) is satisfied, Δt α and ∀α can be 6.1, where the discrepancy between the Pareto optimal fronts for
safely excluded from the set of DVs listed in Table 1 and they can UFjΔt;min ∈½0:90; 1:0½ was lower for Y ZnðC2 H5 Þ2 jΔt o 0:40. In conclu-
be assigned their minimum acceptable values, since the optimiza- sion, it is impossible to attain max Y ZnðC2 H5 Þ2 jΔt while preserving
tions of Eqs. (8) and (9) both strive to minimize Δt α . Furthermore, the process robustness, inherent to higher half-cycle average
A. Holmqvist et al. / Chemical Engineering Science 96 (2013) 71–86 81

Fig. 8. Projections of the feasible objective function space and the feasible inequality constraint space for UFjΔt;min ¼ 0:90 and for the set of DVs u ¼ ½Q_ ZnðC2 H5 Þ2 ; Δt ZnðC2 H5 Þ2 , and
the associated half-cycle average substrate exposure dose, 〈δZnðC2 H5 Þ2 〉 (Eq. (11)). ð○Þ Feasible Pareto optimal solution set, P n ∈U; ðÞ Feasible dominated solutions, u∈U; ð◀Þ Ex
situ XRR film thickness measurement data presented previously (Holmqvist et al., 2012).

Table 2
Analysis of the impact of the sample u^ λ ∼N ðunλ ; s2un Þ, with a percentage deviation of
λ

sunλ ¼ ½1:0; 5:0; 10:0ð%Þ in the set unλ (Eq. (22)) on f ¼ ½P jΔt ; Y ZnðC2 H5 Þ2 jΔt † and on
UFjΔt along the Pareto optimal front for ½UFjΔt;min ; N Ω  ¼ ½0:95; 2.

λ1 n
jμ^ f 1 ðu^ λ Þ −f 1 ðu^ λ Þj (%)
n
jμ^ f 2 ðu^ λ Þ −f 2 ðu^ λ Þj (%) ju^ λ ∉Uja
(%)
ju^ λ j

0.1 0.08,0.63,2.23 0.30,2.18,7.98 10.6,26.4,36.1


0.2 0.10,0.74,2.82 0.26,1.74,6.53 8.06,17.1,30.3
0.4 0.14,1.21,4.36 0.25,1.56,5.59 11.2,29.0,38.6
0.5 0.20,1.26,4.83 0.16,0.98,3.59 0.00,1.40,12.2
0.6 0.21,1.51,5.84 0.13,0.81,3.09 0.00,0.00,0.05
0.8 0.27,1.71,6.78 0.05,0.41,1.70 0.00,0.00,0.00
1.0 0.79,3.56,10.7 0.04,0.17,0.52 0.00,0.00,0.00

a
j  j denotes the cardinality of a set.

6.2.2. Sensitivity analysis


Section 4 describes how the sensitivity analysis involves esti-
mating the parameters μ^ X and ΣðXÞ, ^ which are associated with the
approximated pdf, ϕðXÞ with X ¼ ½u; PjΔt ; Y ZnðC2 H5 Þ2 jΔt ; UFjΔt 
(Eqs. (18) and (19)), from the sample, u^ λ ∼N ðunλ ; s2un Þ. Fig. 11 shows
λ
the impact of individual elements of the DV vector, u, on the
Fig. 9. Analysis of the impact of the sample u^ λ ∼N ðunλ ; s2un Þ, with a percentage
λ uncertainty in f ¼ ½P jΔt ; Y ZnðC2 H5 Þ2 jΔt  and UFjΔt , and it shows the
deviation of sunλ ¼ 10:0ð%Þ in the set unλ (Eq. (22)) and the associated weight vector
λ1 ¼ ½0:1; 0:2; 0:4; 0:5; 0:6; 0:8; 1:0† , on f ¼ ½P jΔt ; Y ZnðC2 H5 Þ2 jΔt † along the Pareto opti- projection of the concentric hyperellipsoid contour (Eq. (23)).
mal front, PF n , for ½UFjΔt;min ; N Ω  ¼ ½0:95; 2 with the set of DVs listed in Table 1. The pattern of scatter plots shows if the correlation is high, as
(−▿−) PF n for sunλ ¼ 0:0ð%Þ (cf. Fig. 3); ðÞ Projection of the model response, f ðu^ λ Þ; do also the forms of the estimated equal probability contours,
ðÞ Estimated μ^ f ðu^ λ Þ of Eq. (18); (–) Projection of the hyperellipsoid with a which are elongated along the principal axes of ΣðXÞ, ^ not parallel
significance level α ¼ 5:0ð%Þ (Eq. (23)); (- -) The principal axes are given by the
eigenvectors of Σðf^ ðu^ λ ÞÞ (Eq. (19)). to the figure axes. Table 3 lists the associated correlation coeffi-
^
cients, ϱ^ X in ΣðXÞ, for each DV and case, λ1 ¼ ½0:1; 0:2; 0:4; …; 1:0.
substrate exposure doses, 〈δα 〉, and hence, the process disturbances The absolute magnitudes of the elements of ϱ^ X , show that the
must be more strictly controlled to ensure high performance when temperature, T, and mass flow of the carrier gas, Q_ N2 , are less
targeting this optimization criterion. sensitive than other variables for all optimization specifications
82 A. Holmqvist et al. / Chemical Engineering Science 96 (2013) 71–86

Fig. 10. Analysis of the impact of the sample u^ λ ∼N ðunλ ; s2un Þ, with a percentage deviation of sunλ ¼ 10:0ð%Þ in the set unλ (Eq. (22)) and associated weight vector
λ
λ1 ¼ ½0:1; 0:2; 0:4; 0:5; 0:6; 0:8† , on UFjΔt . (Histogram) Colored bars represent unfeasible solutions, u^ λ ∉U, that violate the algebraic inequality constraint (Eq. (14)); (–)
Cumulative frequency of the distribution.

and cases λ1 ∈0:0; 1:0 considered. In contrast, the pressure at the elaborated and experimentally validated in the first two parts of
outlet, 〈p〉jΓout , was the most critical variable, having remarkably this paper series (Holmqvist et al., 2012, 2013). The MOP involves
high correlation coefficients for λ1 o 0:6. Moreover, the sensitivity simultaneous optimization of the incommensurable objectives of
of pulse sequence parameters, Δt ZnðC2 H5 Þ2 and Q_ ZnðC2 H5 Þ2 , to UFjΔt reactor throughput and overall precursor conversion, subject to a
increased with λ1 as expected (cf. Section 6.1). In addition, set of operational constraints. The constraints include lower
Δt ZnðC2 H5 Þ2 was the most critical variable with respect to f as bounds for the uniformity of cross-substrate film thickness and
λ1 -1:0, and this places stringent demands on the actuation duration of the post-precursor purge. For this reason, a novel
of valves for consistent precursor delivery, in agreement with model-based methodology for rendering large scale constrained
previous studies (Londergan et al., 2002). MOPs for ALD processes was proposed. The non-dominated Pareto
In conclusion, Figs. 9–11 and Tables 2 and 3 make it clear that optimal solution set was obtained by a MOEA developed and
both the uncertainty in the optimization specifications and the tested for this purpose, and reduced the ALD process design space
sensitivity of the associated optimal DVs differ significantly along into a feasible set of design alternatives.
the Pareto optimal front. It has, however, been possible to identify The comprehensive assessment of the Pareto optimal solution
the critical variables that require strict process control, and to set reported here shows how increasingly stringent constraints on
evaluate the robustness of the process, given the assigned pertur- the uniformity of cross-substrate film thickness and the duration
bation in the optimal DVs, sunλ . of the post-precursor purge affect the process objectives. More
stringent constraints on film thickness uniformity lead to higher
half-cycle average substrate exposure doses, provide a driving
7. Concluding remarks force for the ALD gas–surface reactions towards the trailing edge
of the substrate, at the expense of lowered overall precursor yield.
This paper presents quasi-steady-state, constrained MO of the Moreover, increasing the number of reactor chamber volumes
mechanistic model of the continuous cross-flow reactor system purged during the post-precursor purge period shows that the
F-120 manufactured by ASM Microchemistry Ltd. (Suntola, 1992) productivity criterion is favored more by a reduction in the purge
with temporal precursor pulsing. The process model has been period than by promoting the ALD gas–surface reaction rates.
A. Holmqvist et al. / Chemical Engineering Science 96 (2013) 71–86 83

Fig. 11. Analysis of the impact of the individual elements of the sample u^ λ ∼N ðunλ ; s2un Þ, with a percentage deviation of sunλ ¼ 10:0ð%Þ in the set unλ (Eq. (22)) and associated
λ
weight λ1 ¼ 0:1, on f ¼ ½P jΔt ; Y ZnðC2 H5 Þ2 jΔt † and UFjΔt for the set of DVs listed in Table 1. ðÞ Projection of the model response; (–) Projection of the hyperellipsoid with a
percentage significance level α ¼ 5:0ð%Þ (Eq. (23)); (- -) The principal axes are given by the eigenvectors of ΣðXÞ ^ (Eq. (19)).

The subsequent UA/SA based on LHS to the non-dominated Finally, the contribution of this paper is the novel model-
solutions along the Pareto optimal front was carried out to assess based methodology that considers MO in conjunction with UA/SA,
the impact of variability in the associated set of optimal DVs on the for optimizing manufacturing performance metrics of throughput
optimization specifications and to identify the most critical vari- and precursor utilization in continuous cross-flow ALD reactor
ables. The UA/SA shows that the inherent process robustness is designs without compromising the film quality and process
progressively lost as the precursor yield increases. The process control. The fundamental process understanding gained here is of
objectives and constraints, however, may not necessarily be practical interest when scaling ALD processes to large-surface-
affected by the same DVs along the Pareto optimal front, but the area substrates in high-throughput productions where even small
model-based stochastic methodology presented here allows to variations in the growth rate can be detrimental to the end-use
identify the key DVs. properties.
84 A. Holmqvist et al. / Chemical Engineering Science 96 (2013) 71–86

Table 3 λ normalized weights (–)


Sensitivity of the DVs listed in Table 1 with respect to the uncertainty in the μ mean value of the normal distribution (–)
^ in
elements of ½PjΔt ; Y ZnðC2 H5 Þ2 jΔt ; UFjΔt  based on the correlation coefficients, ϱ,
Πα characteristic function of t and Δt α (–)
Eq. (19). The variability of the input originates from a normal distribution with a
percentage deviation of sunλ ¼ 10:0 ð%Þ in the set unλ (Eq. (22)) and the associated
Σ ðj  jÞ covariance matrix (–)
ξℓ surface reaction stoichiometric coefficient (–)
weight vector λ1 ¼ ½0:1; 0:2; 0:4; …; 1:0† .
ρ density (kg m−3)
u jϱ^ u^ ;PjΔt j  102 jϱ^ u^ ;Y ZnðC j j  102 jϱ^ u^ ;UFjΔt j  102 s standard deviation of the normal distribution (–)
2 H5 Þ2 Δt
ϱ correlation coefficient (–)
T 5.73,2.42,5.87 6.78,3.33,4.80 1.62,3.14,8.08 τ residence time (s)
13.2,10.5,9.34 12.2,10.4,8.40 13.7,20.3,14.9, ϕ empirical probability density function (–)
Q_ N2 2.34,1.51,4.28 7.38,2.38,5.71 5.31,3.27,7.83
ψα molar fraction of gaseous species (–)
8.08,8.27,7.95 11.3,9.65,9.88 12.9,11.8,16.1
Ω spatial domain (–)
〈p〉jΓ out 72.2,61.9,54.9 71.2,58.7,49.6 63.8,56.4,55.2 ωα mass fraction of gaseous species (–)
5.54,12.6,13.0 1.74,12.7,14.1 48.5,46.8,40.0
Q_ Zn ðC2 H5 Þ2 11.7,13.2,11.8 12.9,19.9,24.9 33.9,35.5,36.7 Subscripts and superscripts
7.95,7.60,2.57 31.0,30.9,31.1 36.2,23.0,22.1
Δt ZnðC2 H5 Þ2 13.6,6.20,4.30 12.3,29.2,24.4 41.3,40.2,37.7 0 initial value
7.05,12.9,13.8 31.5,33.8,32.0 40.9,26.4,21.1 α; β gaseous species indices
κ surface species index
i objective function index
Nomenclature j decision variable index
k; l algebraic constraint indices
Roman letters ℓ surface reaction index
AΩ cross section area of the reactor chamber (m2) STP state variable at STP
C algebraic constraint (–) sat substrate surface saturation
Dαβ s solid
binary diffusivity ðm2 s−1 Þ
F system of differential algebraic equations (–)
f objective function vector (–)
g response function (–)
hs film height ðÅÞ Acknowledgment
M molecular weight (kg mol−1)
ms film mass per unit area (kg m−2)
This work has been supported by the Swedish Research Council
Q_ α
volumetric flow rate at STP (sccm)
under Grant no. 2006-3738.
−2 −1
N_ α jΓ total mass flux normal to the boundary Γ (kg m s )
n outward unit vector normal to the boundary Γ (–)
PjΔt productivity (ng cm−2 s−1) References
p pressure (Pa)
Pn Pareto optimal solution set (–) Aarik, J., Aidla, A., Kasikov, A., Mändar, H., Rammula, R., Sammelselg, V., 2006.
Influence of carrier gas pressure and flow rate on atomic layer deposition of
PF n Pareto optimal front (–) HfO2 and ZrO2 thin films. Appl. Surf. Sci. 252 (16), 5723–5734.
R universal gas constant (J mol−1 K−1) Aarik, J., Siimon, H., 1994. Characterization of adsorption in flow type atomic layer
rℓ surface reaction rate (mol m−2 s−1) epitaxy reactor. Appl. Surf. Sci. 81 (3), 281–287.
Adomaitis, R.A., 2010. Development of a multiscale model for an atomic layer
T temperature (K) deposition process. J. Cryst. Growth 312 (8), 1449–1452.
t time (s) Adomaitis, R.A., 2011. A ballistic transport and surface reaction model for simulat-
U multi-attribute utility function (–) ing atomic layer deposition processes in high-aspect-ratio nanopores. Chem.
Vapor Deposition 17 (10-12), 353–365.
UFjΔt film thickness uniformity factor (–)
Aelion, V., Cagan, J., Powers, G., 1991. Inducing optimally directed innovative
u decision variables (–) designs from chemical engineering first principles. Comput. Chem. Eng. 15
U feasible decision space (–) (2), 619–627.
v velocity vector (m s−1) Bakke, J.R., Pickrahn, K.L., Brennan, T.P., Bent, S.F., 2011. Nanoengineering and
interfacial engineering of photovoltaics by atomic layer deposition. Nanoscale
w algebraic variables (–) 3, 3482–3508.
x state variables (–) Baunemann, A., 2006. Precursor chemistry of Tantalum and Niobium nitride for
y model output variables (–) MOCVD and ALD applications. Ph.D. Thesis, Ruhr-University, Bochum.
Benyahia, B., Latifi, M.A., Fonteix, C., Pla, F., 2011. Multicriteria dynamic optimization
Y α jΔt precursor yield (–) of an emulsion copolymerization reactor. Comput. Chem. Eng. 35 (12),
z spatial coordinate (m) 2886–2895.
Bird, R.B., Stewart, W.E., Lightfoot, E.N., 1960. Transport Phenomena, 2nd ed. John
Greek letters Wiley & Sons, Inc., New York.
Burton, B.B., Goldstein, D.N., George, S.M., 2009. Atomic Layer Deposition of MgO
β calibration parameter vector (–) Using Bis(ethylcyclopentadienyl)magnesium and H2 O. J. Phys. Chem. C 113 (5),
1939–1946.
Γ partitioned boundary (–) Carcia, P.F., McLean, R.S., Groner, M.D., Dameron, A.A., George, S.M., 2009. Gas
Δt α pulse duration (s) diffusion ultrabarriers on polymer substrates using Al2 O3 atomic layer deposi-
δα substrate exposure dose (Pa s) tion and sin plasma-enhanced chemical vapor deposition. J. Appl. Phys. 106 (2),
023533.
Δ laminar boundary layer thickness (m)
Cheng, T.S., Hsiao, M.C., 2008. Numerical investigations of geometric effects on flow
ζ substrate boundary local coordinate variable (m) and thermal fields in a horizontal CVD reactor. J. Crys. Growth 310 (12),
η dynamic viscosity of the gas mixture (kg m−1 s−1) 3097–3106.
θκ fractional surface coverage of surface species (–) Cleveland, E.R., Henn-Lecordier, L., Rubloff, G.W., 2012. Role of surface intermedi-
ates in enhanced, uniform growth rates of TiO2 atomic layer deposition thin
Λ maximum molar concentration of surface sites films using titanium tetraisopropoxide and ozone. J. Vac. Sci. Technol. A 30 (1),
(mol m−2) 01A150.
A. Holmqvist et al. / Chemical Engineering Science 96 (2013) 71–86 85

Coello, C.A.C., Lamont, G.B., Veldhuizen, D.A.V., 2007. Evolutionary Algorithms for for advanced gate stack engineering. In: Timans, P.J., Gusev, E., Roozeboom, F.,
Solving Multi-Objective Problems, 2nd ed. Springer, New York. Ozturk, M.C., Kwong, D.L., et al. (Eds.), Rapid Thermal and Other Short-Time
COMSOL AB, 2008. COMSOL Multiphysics User's Guide, Version 3.5a. COMSOL AB, Technologies III, vol. 11. The Electrochemical Society, pp. 163–175.
Tegnérgatan 23, SE-111 40, Stockholm. Masel, R.I., 1996. Principle of Adsorption and Reaction on Solid Surface, 1st ed. John
Dasgupta, N.P., Trejo, O., Prinz, F.B., 2012. Use of a high-flow diaphragm valve in the Wiley & Sons, Inc., New York.
exhaust line of atomic layer deposition reactors. J. Vac. Sci. & Technol. A 30 (1), Matero, R., Rahtu, A., Ritala, M., Leskelä, M., Sajavaara, T., 2000. Effect of water dose
01A110. on the atomic layer deposition rate of oxide thin films. Thin Solid Films 368 (1),
Davis, M.E., 1984. Numerical Methods and Modeling for Chemical Engineers, 1st ed. 1–7.
John Wiley & Sons, Inc., New York. Maydannik, P.S., Kaariainen, T.O., Cameron, D.C., 2012. Continuous atomic layer
Deb, K., 2001. Multi-Objective Optimization using Evolutionary Algorithms. John deposition: explanation for anomalous growth rate effects. J. Vac. Sci. Technol.
Wiley & Sons, Chichester. A 30 (1), 01A122.
Deb, K., Pratap, A., Agarwal, S., Meyarivan, T., 2002. A fast and elitist multi-objective Miettinen, K., 1998. Nonlinear Multiobjective Optimization, 1st ed. Kluwer Aca-
genetic algorithm: NSGA-II. IEEE Trans. Evolutionary Comput. 6 (2), 182–197. demic Publishers, Boston, MA.
Deminsky, M., Knizhnik, A., Belov, I., Umanskii, S., Rykova, E., Bagatur'yants, A., Mousa, M.B.M., Oldham, C.J., Jur, J.S., Parsons, G.N., 2012. Effect of temperature and
et al., 2004. Mechanism and kinetics of thin zirconium and hafnium oxide film gas velocity on growth per cycle during Al2 O3 and ZnO atomic layer deposition
growth in an ALD reactor. Surf. Sci. 549 (1), 67–86. at atmospheric pressure. J. Vac. Sci. Technol. A 30 (1), 01A155.
Elers, K.E., Blomberg, T., Peussa, M., Aitchison, B., Haukka, S., Marcus, S., 2006. Film Nelson, S.F., Levy, D.H., Tutt, L.W., Burberry, M., 2012. Cycle time effects on growth
uniformity in atomic layer deposition. Chem. Vapor Deposition 12 (1), 13–24. and transistor characteristics of spatial atomic layer deposition of zinc oxide.
Elliott, S.D., 2007. Models for ald and mocvd growth of rare earth oxides. In: J. Vac. Sci. Technol. A 30 (1), 01A154.
Fanciulli, M., Scarel, G. (Eds.), Rare Earth Oxide Thin Films. Springer Berlin Park, H.-S., Min, J.-S., Lim, J.-W., Kang, S.-W., 2000. Theoretical evaluation of film
Heidelberg, New York, pp. 73–86. growth rate during atomic layer epitaxy. Appl. Surf. Sci. 158, 81–91.
Fitzpatrick, P.R., Gibbs, Z.M., George, S.M., 2012. Evaluating operating conditions for Poodt, P., Cameron, D.C., Dickey, E., George, S.M., Kuznetsov, V., Parsons, G.N., et al.,
continuous atmospheric atomic layer deposition using a multiple slit gas source 2012. Spatial atomic layer deposition: a route towards further industrialization
head. J. Vac. Sci. & Technol. A 30 (1), 01A136. of atomic layer deposition. J. Vac. Sci. Technol. A 30 (1), 010802.
George, S.M., 2010. Atomic layer deposition: an overview. Chem. Rev. 110 (1), Price, K.V., 1999. An introduction to differential evolution. In: Corne, D., Dorigo,
111–131. M., Glover, F. (Eds.), New Ideas in Optimization. McGraw-Hill, London,
Gobbert, M.K., Prasad, V., Cale, T.S., 2002. Modeling and simulation of atomic layer pp. 79–108.
deposition at the feature scale. J. Vac. Sci. & Technol. B 20 (3), 1031–1043. Price, K.V., Storn, R.M., Lampinen, J.A., 2005. Differential Evolution: A Practical
Granneman, E., Fischer, P., Pierreux, D., Terhorst, H., Zagwijn, P., 2007. Batch ALD: Approach to Global Optimization, 1st ed. Springer, Berlin.
characteristics, comparison with single wafer ALD, and examples. Surf. Coatings Puurunen, R.L., 2003. Growth per cycle in atomic layer deposition: a theoretical
Technol. 201 (22–23), 8899–8907. model. Chem. Vapor Deposition 9 (5), 249–257.
Groner, M.D., Fabreguette, F.H., Elam, J.W., George, S.M., 2004. Low-temperature Puurunen, R.L., 2004. Analysis of hydroxyl group controlled atomic layer deposition
Al2 O3 atomic layer deposition. Chem. Mater. 16 (4), 639–645. of hafnium dioxide from hafnium tetrachloride and water. J. Appl. Phys. 95 (9),
Hangos, K.M., Cameron, I.T., 2001. Process modelling and model analysis. In: 4777–4786.
Stephanopoulos, G., Perkins, J. (Eds.), Process Systems Engineering, vol. 4. Puurunen, R.L., 2005. Surface chemistry of atomic layer deposition: a case study for
Academic Press, London. the trimethylaluminum/water process. J. Appl. Phys. 97 (12), 121301.
Helton, J.C., Davis, F.J., 2002. Illustration of sampling-based methods for uncertainty Qin, A.K., Huang, V.L., Suganthan, P.N., 2009. Differential evolution algorithm with
and sensitivity analysis. Risk Anal. 22 (3), 591–622. strategy adaptation for global numerical optimization. IEEE Trans. Evolutionary
Helton, J.C., Davis, F.J., 2003. Latin hypercube sampling and the propagation Comput. 13 (2), 398–417.
of uncertainty in analyses of complex systems. Reliab. Eng. Syst. Saf. 81 (1), 23–69. Qin, A.K., Suganthan, P.N., 2005. Self-adaptive differential evolution algorithm for
Helton, J.C., Johnson, J.D., Sallaberry, C.J., Storlie, C.B., 2006. Survey of sampling- numerical optimization. IEEE Congress on Evolutionary Computation (CEC
based methods for uncertainty and sensitivity analysis. Reliab. Eng. Syst. Saf. 91 2005), vol. 2, pp. 1785-1791.
(10–11), 1175–1209. Ritala, M., Niinisto, J., 2009. Industrial applications of atomic layer deposition. ECS
Henn-Lecordier, L., Anderle, M., Robertson, E., Rubloff, G.W., 2011. Impact of Trans. 25 (8), 641–652.
parasitic reactions on wafer-scale uniformity in water-based and ozone-based Schiesser, W.E., 1991. The Numerical Method of Lines: Integration of Partial
atomic layer deposition. J. Vac. Sci. Technol. A 29 (5), 051509. Differential Equations, 1st ed. Academic Press, San Diago.
Holmqvist, A., Törndahl, T., Stenström, S., 2012. A model-based methodology for the Schwaab, M., Lemos, L.P., Pinto, J.C., 2008. Optimum reference temperature for
analysis and design of atomic layer deposition processes—part I: mechanistic reparameterization of the Arrhenius equation. Part 2: problems involving
modelling of continuous flow reactors. Chem. Eng. Sci. 81, 260–272. multiple reparameterizations. Chem. Eng. Sci. 63 (11), 2895–2906.
Holmqvist, A., Törndahl, T., Stenström, S., 2013. A model-based methodology for the Schwaab, M., Pinto, J.C., 2007. Optimum reference temperature for reparameteriza-
analysis and design of atomic layer deposition processes—part II: experimental tion of the Arrhenius equation. Part 1: problems involving one kinetic constant.
validation and mechanistic analysis. Chem. Eng. Sci. 94, 316–329. Chem. Eng. Sci. 62 (10), 2750–2764.
Jur, J.S., Parsons, G.N., 2011. Atomic layer deposition of Al2 O3 and ZnO at atmo- Schwaab, M., Pinto, J.C., 2008. Optimum reparameterization of power function
spheric pressure in a flow tube reactor. ACS Appl. Mater. Interfaces 3 (2), models. Chem. Eng. Sci. 63 (18), 4631–4635.
299–308. Siimon, H., Aarik, J., 1995. Modelling of precursor flow and deposition in atomic
Kim, I.Y., de Weck, O.L., 2005. Adaptive weighted-sum method for bi-objective layer deposition reactor. J. de Physique IV 05 (C5), 245–252.
optimization: Pareto front generation. Struct. Multidisciplinary Optim. 29 (2), Siimon, H., Aarik, J., 1997. Thickness profiles of thin films caused by secondary
149–158. reactions in flow-type atomic layer deposition reactors. J. Phys. D 30 (12),
Knoops, H.C.M., Elam, J.W., Libera, J.A., Kessels, W.M.M., 2011. Surface loss in ozone- 1725–1728.
based atomic layer deposition processes. Chem. Mater. 23 (9), 2381–2387. Sneh, O., Clark-Phelps, R.B., Londergan, A.R., Winkler, J., Seidel, T.E., 2002. Thin film
Kuse, R., Kundu, M., Yasuda, T., Miyata, N., Toriumi, A., 2003. Effect of precursor atomic layer deposition equipment for semiconductor processing. Thin Solid
concentration in atomic layer deposition of Al2 O3 . J. Appl. Phys. 94 (10), Films 402 (1-2), 248–261.
6411–6416. Storn, R., Price, K., 1997. Differential evolution—a simple and efficient heuristic for
Lankhorst, A.M., Paarhuis, B.D., Terhorst, H.J.C.M., Simons, P.J.P.M., Kleijn, C.R., 2007. global optimization over continuous spaces. J. Global Optim. 11 (4), 341–359.
Transient ALD simulations for a multi-wafer reactor with trenched wafers. Surf. Sundaram, G.M., Bertuch, A., Bhatia, R., Coutu, R., Dalberth, M.J., Deguns, E., et al.,
Coatings Technol. 201 (22-23), 8842–8848. 2010. Large format atomic layer deposition. ECS Trans. 33 (2), 429–440.
Laumanns, M., Thiel, L., Deb, K., Zitzler, E., 2002. Combining convergence and Suntola, T., 1992. Atomic layer epitaxy. Thin Solid Films 216 (1), 84–89.
diversity in evolutionary multi-objective optimization. Evolutionary Comput. The MathWorks, Inc., 2010a. MATLAB R2010b Documentation, Version 7.11.
10 (3), 263–282. The MathWorks, Inc., 3 Apple Hill Drive, Natick, MA 01760-2098.
Lecordier, L.C., 2009. Wafer-scale process and materials optimization in cross-flow The MathWorks, Inc., 2010b. Optimization Toolbox™ User's Guide, Version 7.11.
atomic layer deposition. Ph.D. Thesis, University of Maryland, College Park, The MathWorks, Inc., 3 Apple Hill Drive, Natick, MA 01760-2098.
Maryland 20742. The MathWorks, Inc., 2010c. Statistic Toolbox™ User's Guide, Version 7.11.
Lei, W., Henn-Lecordier, L., Anderle, M., Rubloff, G.W., Barozzi, M., Bersani, M., 2006. The MathWorks, Inc, 3 Apple Hill Drive, Natick, MA 01760-2098.
Real-time observation and optimization of tungsten atomic layer deposition Vardeman, S.B., 1994. Statistics for Engineering Problem Solving, 1st ed. PWS
process cycle. J. Vac. Sci. Technol. B 24 (2), 780–789. Publishing Company, Boston.
Levy, D.H., Nelson, S.F., Freeman, D., 2009. Oxide electronics by spatial atomic layer Veldhuizen, D.A.V., Lamont, G.B., 2000. Multiobjective evolutionary algorithms:
deposition. J. Disp. Technol. 5 (12), 484–494. analyzing the state-of-the-art. Evolutionary Comput. 8 (2), 125–147.
Lim, J.-W., Park, H.-S., Kang, S.-W., 2001. Kinetic modeling of film growth rate in Wang, Y.-N., Wu, L.-H., Yuan, X.-F., 2010. Multi-objective self-adaptive differential
atomic layer deposition. J. Electrochem. Soc. 148 (6), C403–C408. evolution with elitist archive and crowding entropy-based diversity measure.
Lim, J.-W., Park, J.-S., Kang, S.-W., 2000. Kinetic modeling of film growth rates of TiN Soft Comput. 14 (3), 193–209.
films in atomic layer deposition. J. Appl. Phys. 87 (9), 4632–4634. Wolden, C.A., Kurtin, J., Baxter, J.B., Repins, I., Shaheen, S.E., Torvik, J.T., et al., 2011.
Logist, F., Van Erdeghem, P.M.M., Van Impe, J.F., 2009. Efficient deterministic Photovoltaic manufacturing: present status, future prospects, and research
multiple objective optimal control of (bio)chemical processes. Chem. Eng. Sci. needs. J. Vac. Sci. Technol. A 29 (3), 030801.
64 (11), 2527–2538. Yanguas-Gil, A., Elam, J.W., 2012. Simple model for atomic layer deposition
Londergan, A.R., Ramanathan, S., Vu, K., Rassiga, S., Hiznay, R., Winkler, J., et al., precursor reaction and transport in a viscous-flow tubular reactor. J. Vac. Sci.
2002. Process optimization in atomic layer deposition of high-k oxides Technol. A 30 (1), 01A159.
86 A. Holmqvist et al. / Chemical Engineering Science 96 (2013) 71–86

Yee, A.K.Y., Ray, A.K., Rangaiah, G.P., 2003. Multiobjective optimization of an Zienkiewicz, O.C., Taylor, R.L., 2000. The Finite Element Method—Fluid Dynamics,
industrial styrene reactor. Comput. Chem. Eng. 27 (1), 111–130. 5th ed., vol. 3. Butterworth-Heinemann, Oxford.
Yim, S.-S., Lee, D.-J., Kim, K.-S., Kim, S.-H., Yoon, T.-S., Kim, K.-B., 2008. Nucleation Zienkiewicz, O.C., Taylor, R.L., 2000. The Finite Element Method—The Basis, 5th ed.,
kinetics of Ru on silicon oxide and silicon nitride surfaces deposited by atomic vol. 1. Butterworth-Heinemann, Oxford.
layer deposition. J. Appl. Phys. 103 (11), 113509. Zitzler, E., Laumanns, M., Thiele, L., 2001. SPEA2: improving the strength Pareto
Ylilammi, M., 1995. Mass transport in atomic layer deposition carrier gas reactors. evolutionary algorithm. Technical Report TIK-Report 103, Computer engineer-
J. Electrochem. Soc. 142 (7), 2474–2479. ing and networks laboratory (TIK), Swiss Federal Institute of Technology (ETH),
Yousfi, E.B., Fouache, J., Lincot, D., 2000. Study of atomic layer epitaxy of zinc oxide Zurich.
by in situ quartz crystal microgravimetry. Appl. Surf. Sci. 153 (4), 223–234.

You might also like