Chemical Engineering Science 94 (2013) 316–329

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 II:
Experimental validation and mechanistic analysis
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

c Methodology for estimating kinetic parameters involved in the ALD gas–surface reactions.
c Assessment of the statistical reliability of the parameter estimates and the model response.
c Continuous flow ALD reactor model was validated experimentally and showed high predictability.
c Mechanistic dependence of process operating parameters on ALD film growth profiles were assessed.
c Application of a heuristic optimization algorithm implemented in a cluster environment.

a r t i c l e i n f o a b s t r a c t

Article history: This paper demonstrates the experimental validation and mechanistic analysis of the continuous cross-
Received 20 February 2012 flow atomic layer deposition (ALD) reactor model developed in the first article of this series (Holmqvist
Accepted 28 June 2012 et al., in press). A general nonlinear parameter estimation problem was formulated to identify the
Available online 11 July 2012
kinetic parameters involved in the developed ALD gas–surface reaction mechanism, governing ZnO film
Keywords: growth, from ex situ film thickness measurements. The presented methodology for comprehensive
Atomic layer deposition model assessment considers the statistical uncertainty of least-squares estimates and its ultimate
Experimental model validation impact on the model predicted response. Joint inference regions were determined to assess the
Parameter identification significance of parameter estimates and results indicate that all estimates involved in the precursor
Optimisation
half-reactions were adequately determined. The reparameterization of the Arrhenius equation
Dynamic simulation
effectively decreased the characteristically high correlations between Arrhenius parameters, leading
Kinetics
to improvement in precision of individual parameter estimates. Model predictions of the spatially
dependent film thickness profile with narrow confidence band were in good agreement with both
calibration and validation experimental data, respectively, under a wide range of operating conditions.
The subsequent extensive theoretical analysis exhibits that the experimentally validated model
successfully reproduces the detailed process dynamics revealed by in situ quartz crystal microbalance
and quadrupole mass spectroscopy diagnostics, and thereby provides a supplementary analysis tool.
Finally, the univariate sensitivity analysis revealed the mechanistic dependence of all the measured
process operating parameters on the spatially dependent film thickness profile, resolved at the level of
a single pulse sequence. Hence, the presented model-based framework serves as a means to guide
future research efforts in the field of ALD process optimization.
& 2012 Elsevier Ltd. All rights reserved.

1. Introduction requirement for the understanding and modeling of atomic layer

deposition (ALD), see part I of this article series (Holmqvist et al., in
Knowledge of the microscopic physicochemical growth process press or Elliott, 2007). According to the International Technology
and the underlying macroscopic mass transport is a fundamental Roadmap for Semiconductors (ITRS) (Semiconductor Industry
Association, 2011), process models have the potential to yield
insight that will enable improved efficiency and throughput in
n
Corresponding author. Tel.: þ46 46 222 8301; fax: þ46 46 222 4526. manufacturing and in the design and evaluation of reactors. Due to
E-mail address: [email protected] (A. Holmqvist). the lack of reliable fundamental physical and chemical data for new

0009-2509/$ - see front matter & 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.ces.2012.06.063
 A. Holmqvist et al. / Chemical Engineering Science 94 (2013) 316–329 317

precursors, and the introduction of new process configurations and reactions and can thus be used to estimate the kinetic parameters
regimes, calibration and experimental validation of process models (Holmqvist et al., in press):
have been identified as an important step in the field.
b ¼ ½L,kT ref,i ,i ,Ei y 8i A f1; 2, . . . ,4g ð1Þ
There are two main ways of providing additional physical
Nb
insight into the nature of the reaction mechanisms concerned where b A R is the Nb -dimensional parameter space. However,
and determining reliable estimates of the kinetic parameters. identifying the restricted set of kinetic parameters, b, in an ALD
(i) Ab initio quantum-chemical and quantum-kinetic calculations process model represents a non-trivial task. The two main reasons
(Deminsky et al., 2004; Xu and Ye, 2010), and (ii) model calibra- are:
tion and experimental validation (Lim et al., 2000; Park et al.,
2000). However, the main objective of the published work was to (i) There is a complex interdependence between the sequential
determine the principal features of the kinetic mechanism, and and parallel elementary surface reactions (Pinto et al., 2011)
detailed fluid dynamics was neglected at that stage. It was involved in ALD, such that the reactivity in one half-cycle is
emphasized by Puurunen and Vandervorst (2004) that model influenced by that in the half-cycle preceding it (Kuse et al.,
calibration utilizing experimental saturation curves determined 2003; Ritala and Leskelä, 2002). Consequently, the kinetics of
from flow-type reactors, which are dominated by mass transport each reaction step must be balanced with respect to the other.
(Aarik and Siimon, 1994; Siimon and Aarik, 1995; Ylilammi, 1995), (ii) The intrinsic mathematical structure of the nonlinear Arrhe-
are generally not suitable for extracting kinetic parameters gov- nius equation introduces high correlation between the fre-
erning ALD growth without a reactor scale model. For this reason, quency factor, Ai, and the activation energy, Ei (Buzzi-Ferraris
the novel mathematical model of the continuous cross-flow and Manenti, 2009; Héberger et al., 1987; Nagy and Turányi,
reactor system F-120 by ASM Microchemistry Ltd. (Suntola, 2011; Schwaab and Pinto, 2007, 2008; Schwaab et al., 2008b;
1992) presented in the first article of this series (Holmqvist Varga et al., 2011).
et al., in press) was derived to mechanistically describe the
complex interactions between the gas-phase fluid dynamics, the For this purpose, a carefully chosen experimental design have
mass transport of individual species, and the ALD heterogeneous been recognized as an important element in comprehending the
surface reaction mechanism (Ritala and Leskelä, 2002). Thereby, coupled nature of the ALD growth kinetics and to obtain ade-
the model incorporates all relevant physicochemical phenomena quately determined parameters (Park et al., 2000). The aim of the
governing ALD growth required to assist in parameter estimation. experimental design is to yield the most possible information, in a
Furthermore, in contrast to the mathematical models proposed in statistical sense, for use in parameter estimation and model
the aforementioned cited studies this model, comprises a system validation (Franceschini and Macchietto, 2008). Further, given a
of nonlinear partial differential equations (PDEs), provides spatial set of discriminating experiments (Alberton et al., 2010, 2011), it
resolution of the substrate film thickness profile (Siimon and is important to utilize the experimental information in an optimal
Aarik, 1995, 1997). Given the novel application, the detailed preserving manner. As emphasized by Schwaab et al. (2008a), the
experimental investigation on growth of ZnO films from Zn(C2H5)2 error structure of the kinetic constants in nonlinear models may
and H2O reported in Holmqvist et al. (in press) was performed in be non-convex, open and constitute of discontinuous regions. For
order to assess the validity of the derived mathematical model on this reason, it was recommended that the entire calibration
a discrete set of sampling positions on the substrate, see e.g. parameter vector, b, should be calibrated simultaneously using
Cleveland et al. (2012) and Henn-Lecordier et al. (2011). the entire experimental calibration data set, y, ^ presented in
The overall objective of this study was to experimentally Holmqvist et al. (in press).
validate and demonstrate the generally applicable mechanistic
model for the analysis and design of ALD processes developed in 2.1. Formulation of the nonlinear parameter estimation problem
the first article of this series (Holmqvist et al., in press). This paper
focuses around three main points. (i) To formulate and solve a The nonlinear parameter estimation problem is formulated as
nonlinear parameter estimation problem using ex situ experimen- a general nonlinear optimization problem:
tal data and the modeling environment presented in Holmqvist min FðbÞ
et al. (in press), and (ii) to assess the statistical reliability of the bAR
Nb

 
parameter estimates and the predicted model response, and (iii) to dx
perform an extensive analysis of the experimentally validated ALD s:t: 0¼F ,x,u, b,w,t , xðt 0 Þ ¼ x0
dt
process model dynamics, and to mechanistically determine the y ¼ gðx,u, b,wÞ
sensitivity of the equipment-scale process operating parameters bmin r b r bmax
on film growth. Although assessed in several experimental studies,
with b ¼ ½L,kT ref,i ,i ,Ei y 8i A f1; 2, . . . ,4g
see e.g. Aarik et al. (2006), Jur and Parsons (2011), and Lei et al.
y
(2006), the mechanistic details of the process operating para- xE ¼ ½p,v, oa , yk ,ms  8a, k
 
meters have not yet been thoroughly studied theoretically. yE ¼ ½hs ðzÞ9f^ ,UF y  y ^f ¼ 1 , 3 , 5  9G 9
sub
The reminder of this paper is organized as follows. The non- 6 6 6
linear parameter estimation problem is formulated in Section 2, uE ¼ ½Dt ZnðC2 H5 Þ2 , Dt H2 O ,Ty
while the subsequent section describes the statistical model wE ¼ ½Dt N , Q_ , Q_ , Q_ y
2 N2 ZnðC2 H5 Þ2 H2 O ,/pS9Gout  ð2Þ
validation methodology. Section 4 outlines the optimization envir-
onment. Section 5 presents the results from the validation meth- where b is subject to lower and upper bounds acting as inequa-
odology followed by an extensive mechanistic analysis of the ALD lity constraints and estimated by minimizing an objective func-
process model. Finally, concluding remarks are drawn in Section 6. tion FðbÞ, based on the deviations between the observed, y, ^ and
predicted system response, y. By using the method of lines (Davis,
1984; Schiesser, 1991) and finite element method (Zienkiewicz
s
2. Nonlinear parameter estimation and Taylor, 2000a,b) framework in COMSOL Multiphysics
(COMSOL AB, 2008), the associated system of nonlinear differ-
The inverse method used in this study utilizes the fact that the ential algebraic equations (DAEs), in which x, u and w represent
height of the film contains information on the elementary surface dependent states, free design variables, and algebraic variables, is
318 A. Holmqvist et al. / Chemical Engineering Science 94 (2013) 316–329

denoted by F in Eq. (2). It is noteworthy that response function g, 3.2. Assessment of the accuracy and reliability of the predicted
which transforms and selects those state variables that are model response
experimentally measured, is a nonlinear function of b for all
states x. The weighted sum of squared residuals is used to The net effect of the statistical assumptions that are inherently
quantify the estimation, and is defined as made when applying the least-squares estimates and the propa-
X gation of parametric uncertainty is to produce model predictions
FðbÞ ¼ ½y^ E yE ðxE ,uE , b,wE Þy W1 ^
E ½y E yE ðxE ,uE , b,wE Þ ð3Þ with an uncertainty bound (Draper, 1995). The approximate
E¼1
100ð1aÞ% confidence bands for the response function (Bates
where W is the square weight matrix corresponding to the and Watts, 1988) are given by the collection of all curves
different weights used to normalize the contribution of each yðx,u, b,wÞ belonging to the set:
particular observation E within the N E -dimensional sample space. n
The details underlying the statistical assumptions and their yðx,u, b,wÞ : 9fyðx,u, b,wÞgfyðx,u, b^ ,wÞg9
implications, which lead to the use of the least-squares estimates o
encompass several different aspects of the regression model, are r ½Jðb^ Þy Rðb^ ÞJðb^ Þ1=2 ðNb F ðNb , j;aÞ Þ1=2 ð7Þ
presented elsewhere (Bates and Watts, 1988).
where Jðb^ Þ ¼ @yðx,u, b,wÞð@bÞ1 9b ¼ b^ denotes the parameter Jaco-
bian matrix (Chappell and Gunn, 1998).
3. Statistical model validation via nonlinear parameter
estimation

In accordance with the methodology employed by Pinto et al. 4. Optimization environment

(2011), the statistical model validation methodology developed
for this study is subdivided into two main parts: (i) parameter The minimization of the objective function, FðbÞ, in nonlinear
identifiability and the uncertainty of the estimates and (ii) good- parameter estimation problems is associated with complex numer-
ness-of-fit of the model response and the underlying assumptions ical problems (Buzzi-Ferraris and Manenti, 2009; Moles et al., 2003).
regarding the disturbances. Therefore, the following measures must be taken into consideration
in the formulation of the method: (i) the size of the parameter space,
(ii) the multimodal nature of the objective function, (iii) the
3.1. Assessment of the accuracy and reliability of the parameter continuity of the objective function, and (iv) the sensitivity of the
estimates objective function to each of the model parameters (Hibbert, 1993).
One way to overcome the non-convexity of the optimization
Parameter precision is equivalent in mathematical terms to problem is to consider a method of heuristic optimization, such as
the size of the inference regions of the parameter estimates evolution strategies and genetic algorithms (Price et al., 2005). These
(Hamilton et al., 1982). Based on the assumptions underlying algorithms overcome the starting-point problem resulting from the
least squares estimation, see e.g. Bates and Watts (1988), second- multi-modal nature of the objective function, by sampling the
order statistics (Witkowski and Allen, 1993) were explored for the objective function at multiple, randomly chosen initial points where
construction of the confidence regions and assessing the uncer- the present parameter bounds define the domain, and rely on a large
tainty of b^ in the parameter space b A RNb . The approximate number of objective function evaluations and a random search
100ð1aÞ% joint confidence region, obtained from a second-order character to assure a high probability of finding the global optimum.
Taylor expansion of the objective function, Fðb^ Þ, around the These characteristics lead to prohibitively long computation
minimum point estimate (Schwaab et al., 2008a), is defined by times for the solution of this kind of problem, making non-
sequential solution approaches necessary. According to the ITRS
fb : ½bb^ y Rðb^ Þ1 ½bb^  r Nb F ðNb , j;aÞ g ð4Þ
initiative, time-critical simulation steps or algorithms should
where F ðNb , j;aÞ is the upper a quantile for Fisher’s F-distribution support parallelization on high-performance computation clus-
s
with j ¼ NE Nb degrees of freedom and Rðb^ Þ is the ðNb  N b Þ ters. Hence, the forward model simulator, COMSOL Multiphysics
parameter variance–covariance matrix (Asprey and Macchietto, (COMSOL AB, 2008) integrated with MATLABs (The MathWorks
2000). This approach also suggests the 100ð1aÞ% marginal Inc, 2010a) was extended for implementation in a cluster envir-
confidence interval for bı (Bard, 1974; Draper and Guttman, onment, following the master-slave paradigm (Coello et al., 2007).
1995), ıA f1; 2, . . . ,Nb g: The model-based methodology supports the working procedure as
follows: First, the parameter space, RNb , is sampled with Latin
fb : 9bı b^ ı 9r Sıı ðb^ Þ1=2 t ðj;1a=2Þ g ð5Þ hypercube sampling (LHS) (McKay et al., 1979), using lhsdesign in
TM
the Statistics Toolbox for MATLABs (The MathWorks Inc, 2010c) to
where t ðj;1a=2Þ is the upper a=2 quantile for Student’s t-distribu- determine a feasible input for the heuristic optimization method.
tion and Sıı ðb^ Þ is the ðı,ıÞ th element of Rðb^ Þ. Nevertheless, as Subsequently, the evolution strategy differential evolution (DE) DE/
emphasized by Bates and Watts (1988), Watts (1994) and rand-to-best/2/bin (Storn and Price, 1997; Price, 1999; Price et al.,
Witkowski and Allen (1993), this linearization approach is only 2005) was used to find the optimal least-squares estimates. Applying
approximately applicable to nonlinear models, and in the case of the concept of distributed computation resulted in a scalable and
high nonlinearity it can be quite inaccurate. For this purpose, the affordable cluster solution for the use of both the LHS and DE
likelihood 100ð1aÞ% joint confidence region (Beale, 1960) is algorithms.
defined for all values of b such that Nonetheless, gradient methods outperform direct search
fb : FðbÞFðb^ Þ rs2 Nb F ðNb , j;aÞ g ð6Þ methods both in reliability and speed of convergence (Bard,
1974). Additionally, DE possesses no strict convergence criteria.
where s2 ¼ Fðb^ Þj1 is the variance estimate of s2 . In contrast to Therefore, in order to ensure reliability and to further enhance
Eq. (4), the confidence regions obtained with this method can be convergence of the output optimal calibration parameter vector
disjointed and unbounded, depending on the complexity of the bn A P from DE with respect to the feasible region; in which the
parameter nonlinearity. parameters are linearly independent (Asprey and Macchietto,
 A. Holmqvist et al. / Chemical Engineering Science 94 (2013) 316–329 319

2000; Vajda et al., 1989): the hybrid numerical procedure described in Section 4. The
Nb Nb experimental calibration data sets, y^ E , 8E 2
= f2; 11,23g used for
P ¼ fb A R n
: rank Jðb Þ ¼ Nb 4:(b A R , UðbÞ r Uðb Þg
n n
ð8Þ
evaluation of FðbÞ are reported in Holmqvist et al. (in press)
the gradient-based algorithm lsqcurvefit in the Optimization and the operating conditions associated with the E th index are
TM
Toolbox for MATLABs (The MathWorks Inc, 2010b) was uti- listed in Table 1. The remaining data sets are used for experi-
lized to yield the final least-squares estimates, b^ . mental model validation (see Section 5.1.3).

5.1.1. Assessment of the accuracy and reliability of the parameter

5. Results and discussion
estimates
The statistical significance and reliability of b^ is assessed by
5.1. Parameter estimation and statistical model validation
evaluating the confidence statements presented in Section 3.1.
A large confidence interval (Eq. (5)) or region (Eqs. (4) and (6))
Parameter estimation of the kinetic parameters (Eq. (1))
suggests either that a parameter estimate cannot be discriminated
involved in the ALD gas–surface reactions was performed with
from another with the experimental design considered or that there
Table 1 is a high uncertainty in the precision associated to this parameter,
Outline of the experimental design. The ALD pulse sequence considered for the growth i.e. the model is inadequate in distinguishing a phenomenon in the
of ZnO was the following ðDt ZnðC2 H5 Þ2 , Dt N2 , Dt H2 O , Dt N2 Þ including a nitrogen purge system (Bard, 1974; Bates and Watts, 1988).
DtN2 ¼ 4:0  101 s between the non-overlapping alternate injections of the precur- The approximate marginal confidence intervals (Eq. (5)) and the
sors for all cases E. All experiments were performed with the process operating corresponding margin of error, eb^ ¼ Sıı ðb^ Þ1=2 t ðj;1a=2Þ , with a sig-
parameters: /pS9Gout ¼ 2:0  102 Pa, ½Q_ ZnðC2 H5 Þ2 , Q_ H2 O , Q_ N2  ¼ ½0:11,0:16,5:0  102 ı
nificance level of a ¼ 5:0  102 and 8ıA N b are reported in Table 2.
sccm [standard cubic centimeters per minute at STP] (cf. Holmqvist et al., in press
The narrow confidence intervals for parameters b^ ı and
for physical interpretation). The uniformity factors for the experimentally observed
response, UF y^ , and the model generated response, UFy, and the relative model error, dy,
8ıA f1; 2, . . . ,5g indicate that these estimates are adequately deter-
for all calibration sets are also presented. mined. Consistently larger relative errors, though still remaining
moderate, were obtained for parameters b^ ı and 8ıA f6; 7, . . . ,9g, i.e.
E DtZnðC2 H5 Þ2  101 DtH2 O  101 T  10  2 UF y^ UFy dy the kinetic parameters governing the rehydroxylation reaction (see
(s) (s) (1C) (%) (%) (%)
e.g. Holmqvist et al., in press; Deminsky et al., 2004; Matero et al.,
2000; Rahtu et al., 2001), and these are thus the most difficult
1 2.0 4.0 1.0 97.1 94.4 5.9
2a 4.0 4.0 1.0 94.7 95.6 4.1 parameters to statistically discriminate against others satisfactorily
3 6.0 4.0 1.0 93.8 95.9 6.7 with the experimental design considered.
4 12.0 4.0 1.0 90.1 96.4 5.7 In addition, the maximum molar concentration of surface sites
5 18.0 4.0 1.0 94.7 95.9 2.7 L can be deduced from the physical properties determined from
6 6.0 1.0 1.0 92.3 96.2 9.9
the analysis of film growth experiments (Coltrin et al., 2008). The
7 6.0 2.0 1.0 93.7 95.8 9.3
8 6.0 4.0 1.0 93.8 95.6 6.1 experimental surface site density, determined from the density of
deposited ZnO films, rs ¼ 5:4  103 kg m3 at T ¼423.2 K in
9 1.0 4.0 1.5 96.8 96.9 2.2 ~ 2=3 N~ ¼
10 2.0 4.0 1.5 95.7 97.8 5.3 Törndahl et al. (2007), was found to be L ¼ ðrs M 1 s NÞ
5 2
11a 4.0 4.0 1.5 98.4 99.1 1.9 1:94  10 mol m . This value shows reasonable agreement
12 6.0 4.0 1.5 98.3 99.4 2.2 with the final estimate b^ 1 presented in Table 2, further verifying
13 12.0 4.0 1.5 99.0 99.9 0.76 that this estimate was adequately determined.
14 6.0 1.0 1.5 97.8 95.8 3.0
Another aspect affecting the reliability of b^ is parameter
15 6.0 2.0 1.5 95.7 99.5 7.3
16 6.0 4.0 1.5 98.4 99.4 2.2 correlation, which may originate from an inappropriate model
representation and/or experimental design (Schwaab and Pinto,
17 1.0 4.0 2.0 97.7 97.9 2.8
18 2.0 4.0 2.0 97.5 98.4 2.9 2007). Despite the implication of the intrinsic mathematical
19 4.0 4.0 2.0 96.8 99.1 5.9 structure of the Arrhenius equation (Nagy and Turányi, 2011;
20 6.0 4.0 2.0 98.9 99.0 3.8 Varga et al., 2011), the majority of the elements of the parameter
21 12.0 4.0 2.0 99.1 99.6 5.1 correlation matrix, C ı‘ ðb^ Þ and 8ıa ‘ A Nb , are moderately low (see
22 6.0 1.0 2.0 99.0 99.0 1.0 P
23a 6.0 2.0 2.0 99.1 99.6 1.3
Table 2) and yielding an overall correlation level C ¼ ð ı a ‘
24 6.0 4.0 2.0 99.0 99.1 3.8 C ı‘ ðb^ Þ2 ðN2b N b Þ1 Þ1=2 (Pritchard and Bacon, 1978) of 0.43. This
implies that the characteristically high parameter correlations,
a
Experimental validation set. almost equal to one (Héberger et al., 1987), were successfully

Table 2
Regression analysis of the parameter least-squares estimates, b^ . The correlation matrix Cðb^ Þ and normalized margin of error e b^ were determined at a significance level of
a ¼ 5  102 . The specific reaction rate kT ref,i ,i and 8i A f1; 2, . . . ,4g are defined at the reference temperature T ref,i ¼ 4:19  102 K. According to Elam and George (2003), the
variable numbers of surface hydroxyl groups that react with the ZnðC2 H5 Þ2 adsorptive precursor molecule in the elementary surface reactions (cf. Holmqvist et al., in press)
was defined as n ¼ 1:37. Parameter units: kT ref,1 ,1 ððmol m2 Þ1n Pa1 s1 Þ; kT ref,2 ,2 ððmol m2 Þn1 Pa1 s1 Þ; kT ref,3 ,3 ððmol m2 Þ1 s1 Þ; kT ref,4 ,4 ðPa1 s1 Þ.

ı Parameter b Parameter estimates b^ Margin of error e b^ (%) Approximate correlation matrix Cðb^ Þ

1 L 1.53  10  5 0.183 1.00 0.63 0.74 0.46 0.52  0.80 0.61 0.39  0.35
2 kref,1 8.16  10  6 1.24 1.00  0.53 0.12  0.22 0.19 0.11 0.10  0.06
3 E1 3.57  104 0.172 1.00 0.21  0.15  0.72 0.19 0.21  0.23
4 kref,2 1.33  10  6 1.16 1.00 0.52  0.28 0.13  0.97  0.30
5 E2 4.04  104 0.155 1.00  0.11 0.09 0.20  0.17
6 kref,3 3.45  10  6 3.29 1.00  0.56  0.31 0.36
7 E3 2.67  104 2.56 1.00 0.38 0.29
8 kref,4 4.68  10  7 5.21 1.00  0.59
9 E4 3.79  104 4.95 1.00
320 A. Holmqvist et al. / Chemical Engineering Science 94 (2013) 316–329

decreased through reparameterization of the Arrhenius equation

and definition of a reference temperature, see Holmqvist et al. (in
press), Schwaab et al. (2008b), and Schwaab and Pinto (2007,
2008), leading to significant improvement in the precision of
individual parameter estimates. In this study, T ref,i , 8i, were
assigned the inverse average temperature over the experimental
range analyzed (cf. Table 1).
In order to understand the behavior of the relative errors and
the parameter correlations, a more rigorous statistical evaluation
of the parameter estimates is obtained by considering confidence
regions (Eqs. (4) and (6)), described by function evaluations
performed by the heuristic optimization method along the search,
as discussed in Schwaab et al. (2008a). The likelihood 100ð1aÞ%
joint confidence regions, with a significance level of a ¼ 5  102 ,
for pairs of parameters bı and 8ıA f1; 2,3; 6g are depicted in Fig. 1.
From the figure it is evident that the significant reduction in
parameter correlation, through the reparameterization method
applied in Holmqvist et al. (in press), also provides an improve-
ment in the elliptical representation of the confidence regions
(Bates and Watts, 1988) associated with the pair of parameters
b2 b3 (of the gas–surface reaction i¼1 Holmqvist et al., in press).
Furthermore, it is noteworthy that the pairwise plots of the
likelihood joint confidence regions clearly do not extend into
negative parameter regions, which is required for interpreting
physically meaningful parameters (Beale, 1960).
Fig. 2. Nominal 95% likelihood confidence region for the normalized and centered
However, when considering the 95% likelihood joint confi- parameters b4 and b8 . (– –) Approximate marginal 95% confidence interval
dence region (Eq. (6)) associated with ıA f4; 8g depicted in Fig. 2, ð7 e b^ Þ; (–) approximate joint 95% confidence region; (þ ) Normalized and
ı
it is evident that the approximate 95% joint confidence region centered least-squares estimates, b^ ı .

Fig. 1. Nominal 95% likelihood confidence region for pairs of normalized and centered parameters bı and 8ıA f1; 2,3; 6g; (– –) approximate marginal 95% confidence interval
ð 7 e b^ Þ; (þ ) normalized and centered least-squares estimates, b^ ı .
ı
 A. Holmqvist et al. / Chemical Engineering Science 94 (2013) 316–329 321

(Eq. (4)) provides a poor representation, and thus, lead to erro- uncertainty on the precision of the parameter estimates given by
neous conclusions about the confidence on model parameters. In the model can be evaluated.
Fig. 2, the point estimate of model parameters is placed at the The graphic representation of some of the experimental
center of the hyper-ellipsoid and the likelihood confidence region calibration data sets E A f1; 3,8; 9,10; 14,17; 20,24g in Fig. 3a–c
is non-convex and unbounded, since the parameter b8 does not indicates that the model output and the experimental data show
have an upper limit. Hence, the model is unable to discriminate good agreement. The comparison between the uniformity factors
between high values of b8 , as high values of this parameter lead to for the experimentally observed, UF y^ , and model generated
a constant model response, y. The occurrence of the elongated response, UF y , in Table 1 also shows satisfactory agreement for
likelihood joint confidence region reflects the parameter correla- the majority of the calibration data sets. The worst fit is found for
tion (Emig and Hosten, 1974). In order to further reduce the the calibration data sets employing a temperature of 1:0  102 1C
parameter correlation and simultaneously improve the precision and short pulse duration of the adsorptive precursors, which give
of the parameter estimates, two essential points are: (i) use an film profiles with the lowest uniformity factors. This is evident
experimental design with extended information that enables when visualizing the model accuracy, the associated bandwidth
discrimination (Franceschini and Macchietto, 2008) of the afore- of the inference bands and when analyzing the relative error dyE
mentioned parameters and (ii) perform the numerical optimiza- and 8EA NE for each calibration index (see Table 1). On the other
tion procedure proposed in Schwaab et al. (2008b) and Schwaab hand, the calibration sets with longer adsorptive precursor pulse
and Pinto (2007, 2008) to provide the optimum reference tem- duration, especially for the H2O precursor, have significantly
peratures, T ref,i , for every ith Arrhenius reaction rate equation, lower relative errors, demonstrating that the model describes
respectively. An optimal experimental design and a set of optimal this part of the calibration region better.
reference values holds the potential to yield precise estimates of When growing oxide films by ALD in a viscous flow reactor
uncorrelated parameters, and the likelihood joint confidence with alternating injection of adsorptive precursors, Groner et al.
region would tend towards being elliptical. (2004) reported that complete purging of reactants, especially
Moreover, on the subject of improving information content, in H2O, takes significant longer time at low temperatures ð r1:0 
a statistical sense, obtained from the measured response, y, ^ to 102 1CÞ because of slower desorption from the reactor walls.
maximize the calibration parameter discrimination capacity, the Hence, purge times must be increased with decreasing tempera-
ITRS guidelines promote the use of in situ diagnostics, including ture in order to prevent simultaneous presence of both precursors
quartz crystal microbalance (QCM) (Aarik et al., 2006; Elam and in the gas phase which results in regular CVD growth, comprising
George, 2003) and quadrupole mass spectroscopy (QMS) (Lei a higher growth rate than in true ALD (Elers et al., 2006; Jur and
et al., 2006; Rahtu and Ritala, 2002). These monitoring techniques Parsons, 2011). In this study, the N2 purge time was held constant
provide extensive dynamic information during the ALD pulse at Dt N2 ¼ 4:0  101 s (see Table 1) in the entire temperature
sequence, thereby improving the information content and enhan- region and local CVD growth may occur on the leading edge of the
cing the reliability of the parameter estimates. However, the main substrate in the cross-flow reactor, causing the sharp change in
objective of the derived ALD process model in this article series growth rate over the substrate, and thus degrading the overall
was to accurately predict the response of the substrate spatially uniformity. Consequently, this phenomena could explain the
dependent film thickness profile to the variations of all the discrepancy between the model predicted response and the
measured equipment-scale process operating parameters. For this experimental measurement data at T ¼ 1:0  102 1C.
reason, the ex situ instrumentation used for film characterization A more comprehensive presentation of the reliability and the
in Holmqvist et al. (in press) are superior to the aforementioned adequacy of the model response can be achieved by examining
in situ diagnostics, since the latter measurement response pro- the distribution of the residual errors for the calibrated model.
vides no information regarding the nature of the spatial depen- This can be graphically visualized by plotting the relative error dyE
dence of the deposition rate. Consequently, to further enhance the and 8EA N E for all calibration sets against the cumulative distribu-
reliability of the parameter estimates, b^ , while still maintaining tion, determined using the empirical cumulative distribution
TM
the precision of the spatial model response, the utilization of both function ecdf in the Statistics Toolbox for MATLABs (The
in situ and ex situ diagnostic methods may become necessary. MathWorks Inc, 2010c). On the basis of the cumulative distribu-
tion of residual errors in Fig. 4, it is evident that 90% of the
residual errors is below 7.3% and the only errors above this arise
5.1.2. Assessment of the accuracy and reliability of the predicted from experimental indices E A f6; 7g. However, the relative errors
model response are below 10% for all experimental indices, confirming rather high
In order to verify the precision of the model response, the predictability even for the worst cases studied.
statistical uncertainties and their effects on the model response The fitted normal cumulative distribution function, using
TM
were addressed by considering the model 100ð1aÞ% approximate normcdf in the Statistics Toolbox for MATLABs (The
confidence band (Eq. (7)). The model generated spatially dependent MathWorks Inc, 2010c), depicted in Fig. 4, demonstrates reason-
film thickness profile, hs ðzÞ, for the least-squares estimates (see able agreement with the cumulative frequency of the relative
Table 2) and the associated 95% approximate confidence band are model error, implying that the residual errors follow a normal
depicted in Fig. 3 along with the experimentally observed film distribution with an expected value and a variance (Hines et al.,
thickness, y9 ^ f^ , sampled at three spatially equidistant positions, 2003), in accordance with the chi-square goodness-of-fit test as
f^ ¼ ½16 , 36 , 56  9Gsub 9 (m), on the substrate, Gsub . defined in Schwaab et al. (2006), Pinto et al. (2011), and Alberton
Essentially, to statistically assess the reliability of the spatially et al. (2011). Thus, the underlying statistical assumptions on
dependent film thickness profile hs ðzÞ, it is necessary that the which the use of the least squares was based are not violated.
confidence bandwidth is moderately narrow and that it replicates
the spatial dependency of the model response. According to Fig. 3
it is possible to statistically assure that the model output predicts 5.1.3. Experimental model validation
a non-uniform film thickness profile for E ¼ 17, in contrast to the The ultimate test of the adequacy of the model is to compare it
case of E ¼ 10. Hence, the accuracy and predictability of the model to the validation set, and if the experimentally observed and
can thus be determined by analyzing the bandwidth and shape of model predicted response show good agreement, the model can
the approximate inference band. In addition, the effect of the be considered valid. The region of validity can subsequently be
322 A. Holmqvist et al. / Chemical Engineering Science 94 (2013) 316–329

1.14 1.49 1.48

1.03 1.34 1.34

0.93 1.2 1.2

0.82 1.06 1.07

0.71 0.92 0.93

1.98 2.27 2.22

1.86 2.1 2.06

1.74 1.93 1.9

1.62 1.76 1.74

1.49 1.59 1.58

2.09 2.21 2.21

1.96 2.07 2.07

1.84 1.92 1.92

1.72 1.78 1.78

1.59 1.63 1.64

1.26 2.27 2.01

1.16 2.17 1.93

1.06 2.07 1.86

0.95 1.97 1.79

0.85 1.87 1.72

Fig. 3. Spatially dependent film thickness profile hs ðzÞ as a function of the substrate boundary Gsub local coordinate variable z A ½0,9Gsub 9 (m). The operating conditions
associated with the E th index are specified in Table 1. Model response hs ðzÞ (Å cycle  1) for: (a) calibration data sets performed at T ¼ 1:0  102 1C; (b) calibration data sets
performed at T ¼ 1:5  102 1C; (c) calibration data sets performed at T ¼ 2:0  102 1C; (d) validation data sets. (þ ) Experimental response hs ðf^ Þ (Å cycle  1) at f^ ¼ ½16 , 36 , 56 
9Gsub 9 (m); (–) spatial film thickness profile; (– –) 95% approximate confidence band (Eq. (7)); (––) uncertainty band for the expected response determined by uniformity
distribution with the parameter bounds specified according to the marginal confidence intervals given in Table 2.

defined as the union between of the calibration region and the 5.2. Mechanistic analysis of the ALD process model
region covered by the validation experiments (Brereton, 2003).
Fig. 3d visualizes the model response for the experimental This section demonstrates the universal applicability of the
validation data set with EA f2; 11,23g (see Table 1). The uncertain- mechanistic model of the continuous cross-flow ALD reactor
ties enclosed bands were constructed from the extreme model (Suntola, 1992), derived in the first article of this series
response generated from Latin hypercube sampling (LHS) (McKay (Holmqvist et al., in press), with statistical reliable least-squares
TM
et al., 1979), using lhsdesign in the Statistics Toolbox for estimates (see Eq. (1) and Table 2) for the kinetic parameters
s
MATLAB (The MathWorks Inc, 2010c), with the parameter involved in the elementary ALD gas–surface reactions. Based on
boundaries specified in accordance with the marginal confidence the experimentally validated mathematical model, with adequate
intervals (see Eq. (5) and Table 2). The graphical presentation of predictability of the spatial resolution of the film thickness profile
the model response in Fig. 3d demonstrates that the performance (see Section 5.1), an extensive theoretical investigation was
of the model and the resulting predictability are satisfactory, with performed with the focus on mechanistic understanding of ALD
low relative residual errors, dy, and conformal uniformity factors process dynamics and the influence of equipment-scale process
for the model generated, UF y , and the experimentally observed, operating parameters on ALD film growth. The mechanistic
UF y^ , response, which confirms that the model is valid within the model-based analysis complements available experimental
calibration region. results deduced from the literature, see e.g. Aarik et al. (2006),
 A. Holmqvist et al. / Chemical Engineering Science 94 (2013) 316–329 323

2 1
100
1.8 0.9

1.6 0.8
80
1.4 0.7

1.2 0.6
60
1 0.5

0.8 0.4
40
0.6 0.3

0.4 0.2
20 0.2 0.1

0 0
0 50 100 150 200 250 300
0
0 2 4 6 8 10 12 14
Fig. 5. Dependence of the film mass increment on the process temperature during
one ALD sequence. (–) The integral mean value of the mass increment /ms S over
Fig. 4. The cumulative distribution of the relative residual errors, dyE , and 8E A N E . Gsub per ALD sequence and unit area; ð,Þ analogous to (–) but using the
(–) The cumulative frequency for the calibration data set; (– –) fitted normal generalization of Fick’s law (cf. Holmqvist et al., in press); (– –) the degree of
TM
cumulative distribution function normcdf in the Statistics Toolbox for MATLABs hydroxylation /gS (a.u.) after the H2 O exposure in the pulse sequence. Process
(The MathWorks Inc, 2010c). operating parameters: /pS9Gout ¼ 3:0  102 Pa, /pS9Gin ¼ 4:3  102 Pa, ½Q_ ZnðC2 H5 Þ2 ,
Q_ H2 O , Q_ N2  ¼ ½0:10,0:15,5:0  102 sccm (standard cubic centimeters per minute at
STP), pulse sequence, ½Dt ZnðC2 H5 Þ2 , Dt N2 , Dt H2 O , Dt N2  ¼ ½2:0,4:0,2:0,4:0  101 s.
Elam and George (2003), Jur and Parsons (2011), Kuse et al.
(2003), Lei et al. (2006), and Rahtu et al. (2001).
adsorption sites for the adsorptive H2O precursor in the subse-
quent pulse sequence. The pseudo-steady-state process corre-
5.2.1. General characteristics of the ALD film growth process sponding to one pulse sequence can be seen in the figure, where
In the context of deposition of ZnO films in continuous cross- the initial surface fraction composition is also governed at the end
flow ALD reactors, the temperature, which is the most readily of the same pulse sequence. This is a consequence of the initial
attribute to activation energies of the elementary reactions, as normally hydroxylated surface, resulting in a constant mass gain
well as the process time parameters, i.e. precursor exposure and per cycle (MGPC) (Burton et al., 2009), and consistent with the
purge times, dependence have been undertaken seriously in Elam ex situ XRR analysis in Holmqvist et al. (in press).
et al. (2002), Elam and George (2003), Törndahl et al. (2007), and However, all states governing the system’s behavior do not
Yousfi et al. (2000). According to Deminsky et al. (2004) and remain constant throughout the duration of the pulse sequence;
Matero et al. (2000), either the dehydroxylation reaction or the therefore, the system is evidently not time-invariant. The
rehydroxylation reaction (surface reaction index i A f3; 4g, response of the transient system is apparent when analyzing
Holmqvist et al., in press) determines the overall growth rate in the mean mass deposited per unit area resolved at the level of
the high-temperature region. This is evident when considering each cycle, see Fig. 7. The theoretical trajectory resembles
the temperature dependence of the degree of hydroxylation, /gS, previous experimental QCM measurements for ZnO: i.e. a mass
after the H2O sequence t A ½0, Dt ZnðC2 H5 Þ2 þ Dt N2 þ Dt H2 O , graphically increment due to irreversible adsorption of the adsorptive
presented in Fig. 5, and expressed as Zn(C2H5)2 precursor, a flat portion of the trajectory during the
Z Z purge exposure, and a decrease in mass during the adsorptive
1 yOH
/gS ¼ @G@t ð9Þ H2O precursor sequence (Elam and George, 2003; Yousfi et al.,
9Gsub 9 t Gsub yOH þ yO
2000).
At the highest temperatures, dehydroxylation is so extensive that The general formalism of the sequential elementary surface
it overrides the increase in hydroxyl groups during the water reactions i A f1; 2g (Holmqvist et al., in press) incorporating a
sequence. In contrast, the overall growth rate decreases in the variable number of reacting hydroxyl groups, n, enables the
low-temperature region, which is attributed to the activation surface chemistry to be analyzed during one pulse sequence,
energy of the elementary surface reactions i A f1; 2g. and the discrepancy between different reaction mechanisms to be
A graphic representation of the transient production and investigated on the basis of mass increment during the respective
consumption of fractional surface species, yk and k A fOH, half-reactions. The theoretical mass increment trajectories shown
ZnðC2 H5 Þ2n ,Og, during one complete ALD pulse sequence in Fig. 7 demonstrate the characteristic dependence of n A 0, nL ½,
ðDt ¼ Dt ZnðC2 H5 Þ2 þ 2Dt N2 þ Dt H2 O Þ sampled at three different tem- where nL represents the total number of ligands of the adsorptive
peratures is presented in Fig. 6. It is evident from this figure that organometallic precursor, and corresponding DM 2 DM 1 1 ratios
the fractional surface coverage of elementary oxygen, yO , is the (Elam and George, 2003). Thus, the inherent structure of the
source of the decrease in yOH at elevated temperatures. Because of model supports the utilization of the entire information content
the sequential nature of the surface reactions, the effect of the net of observed mass increment trajectories when extracting detailed
yO production is twofold: the fractional surface coverage of yOH is surface chemistry data, i.e. when determining n and the calibra-
reduced, which limits the number of adsorption sites for the tion parameter vector, b (Eq. (1)).
adsorptive ZnðC2 H5 Þ2 precursor, thereby preventing the produc- An additional method of assessing n considers temporal and
tion of yZnðC2 H5 Þ2n , which successively reduces the number of spatial integration over an arbitrary boundary G, of the reaction
324 A. Holmqvist et al. / Chemical Engineering Science 94 (2013) 316–329

100
1

0.8
80
0.6

0.4
60
0.2

0
40

1
20
0.8

0.6
0
0.4
15
0.2
12
0
9
6
1
3
0.8 0
0.6

0.4 0 0.2 0.4 0.8 1 1.2

0.2

0 Fig. 7. (a) Transient analysis of the integral mean mass increment /ms Sðng cm2 Þ
at Gsub during one ALD sequence. (–) /ms S with the number of reacting hydroxyl
groups n ¼ 1:37; (– –) /ms S with the number of reacting hydroxyl groups n ¼ 1:0.
0 0.2 0.6 0.8 1.2 (b) Transient analysis of the integral mean gas-phase composition, /pa S (Pa), of
the adsorptive precursor at Gin . (–) /pZnðC2 H5 Þ2 S; ð,Þ analogous to (–) but using the
generalization of Fick’s law (cf. Holmqvist et al., in press); (– –) /pH2 O S. The
Fig. 6. Integral mean value of the fractional surface coverage /yk S (a.u.) at Gsub dotted lines perpendicular to the t-axis represent the individual adsorptive
precursors and purge exposures of one pulse sequence. The process operating
during one ALD sequence at: (a) T ¼ 1:5  102 1C; (b) T ¼ 2:0  102 1C;
parameters are analogous to those defined in Fig. 5 and T ¼ 1:5  102 1C.
(c) T ¼ 2:5  102 1C. (-) /yOH S; ð,Þ analogous to (-) but using the generalization
of Fick’s law (cf. Holmqvist et al., in press); (- -) /yZnðC2 H5 Þ2n S; ð  Þ /yO S. The
dotted lines perpendicular to the t-axis represent the individual adsorptive
precursors and purge exposures of one pulse sequence. The process operating Together, Figs. 7 and 8 reveal rigorous process dynamics of the
parameters are analogous to those defined in Fig. 5. ZnO ALD process that are essential process analysis, design and
optimization.
byproduct released during the surface reaction i¼1 (Holmqvist The comparison of the model response in Figs. 5–8, using the
et al., in press) and the entire pulse sequence. Consequently, this binary and multicomponent formulations (cf. Holmqvist et al., in
quantity is governed by press), respectively, for the mass flux show only minor deviations.
This is due to the low molar composition of C2 H6 relative the
Y Z Z ð1ÞW þ 1
remaining species, and hence, the rigorous, multicomponent
n ¼ nL pC2 H6 @G@t ð10Þ
W¼1 tW G formulation may be omitted, to a very good approximation, and
one-way solute–solvent interactions suffice. Therefore, the dis-
where W ¼ f1; 2g and t 1 A ½0, Dt ZnðC2 H5 Þ2 þ Dt N2  for the nominator cussion regarding the sensitivity of the process operating para-
respective t 2 A ½0, Dt ZnðC2 H5 Þ2 þ 2Dt N2 þ Dt H2 O  denominator tem- meters in Section 5.2.2 will be addressed solely on the basis of the
poral integrals. The trajectories of the integral mean value of the binary formulation for the mass flux.
gas-phase composition at the outlet boundary, Gout , during one
pulse sequence are presented graphically in Fig. 8. It is clear that
the /pC2 H6 S9Gout trajectory is explicitly governed by xa,i /r i S9G 5.2.2. Sensitivity of process operating parameters on ALD film
sub
and Eq. (10) provides the specified nominal value of n. growth
The aforementioned method, which considers spatial integra- In their experimental investigation of deposition efficiency in a
tion, provides quantitative, chemically specific information and continuous flow ALD reactor, Aarik et al. (2006) and Jur and
can be used as a supplementary theoretical analysis quantity in Parsons (2011) observed the influence of carrier gas flow para-
in situ QMS diagnostics (Lei et al., 2006; Rahtu et al., 2001). It is meters on ALD growth rate. To verify whether the predicted model
noteworthy that the spatial boundary integration or point evalua- response resemble the experimental observations, a comprehen-
tion can be utilized anywhere downstream of the substrate to sive univariate sensitivity analysis of the measured process oper-
adequately reproduce the output of QMS sampling systems. ating parameters u ¼ ½/pS9Gout ,T, Q_ a , Dt a y , a A fN2 ,ZnðC2 H5 Þ2 ,H2 Og,
 A. Holmqvist et al. / Chemical Engineering Science 94 (2013) 316–329 325

10 However, it is reasonable to exclude the decrease in reactor

1
chamber residence time t ¼ ð9Gsub 9 þ29Gwall 9Þ/vS9O (s) at ele-
vated linear flow rates as the underlying cause of the observed
significant decrease in /ms S, since t 5 Dt a , a A fZnðC2 H5 Þ2 ,H2 Og,
8 for the entire range of /vS9O studied. In contrast, the decrease in
/ms S can be mechanistically correlated to the hydrodynamic
interdependence governed by Eq. (11), which originates from the
6 ideal gas law. At constant applied /pS9Gout and T, Eq. (11) states
that the linear flow rate is directly proportional to the mass flow
of the carrier gas, /vS9O pQ_ N2 . Consequently, the increase
4 in Q_ N2 with /vS9O imposes a decrease in local mass fraction of
the a th adsorptive precursors, oa , and ultimately resulting in
reduced driving force in the gas–surface reaction rates, hence
r i pðpca Þ9G and 8i a 3 (Holmqvist et al., in press).
2 sub
According to Eq. (11), a decrease in carrier gas pressure at Gout ,
at constant applied /vS9O and T, imposes a decrease in Q_ N2 and
an increase in the overall local oa , respectively. Additionally,
0 overall lower local p also implies that the binary gas diffusion
coefficient is increased, since Dab pp1 (Bird et al., 1960), which
10 increases the rates of impingement mass flux to the substrate
surface, ja ¼ rDab roa (Holmqvist et al., in press). In summary,
8
these mechanistic characteristics will promote a higher molar
6 composition of the a th adsorptive precursors, ca , in the vicinity
4 of the substrate, enhancing the rate of elementary surface reac-
tions. In contrast, a decrease in the local carrier gas pressure will
2
limit this driving force. However, the net increase in mass
0 deposited, /ms S, can be mechanistically correlated to the inverse
pressure dependence of the impingement mass flux, ja , which
compensates for the reduction in reaction driving force,
r i pðpca Þ9G for 8i a 3.
sub
0 0.2 0.4 0.8 1 1.2 The temperature dependence on /ms S is thoroughly dis-
cussed in Section 5.2.1. The interdependence of the process
temperature on the applied pressure and linear flow rate of the
Fig. 8. (a) Transient analysis of the integral mean gas-phase composition, /pa S
carrier gas is graphically presented in Fig. 9b and c. It is evident
(Pa), at Gout during one ALD sequence. (–) /pZnðC2 H5 Þ2 S; ð,Þ analogous to (–) but
using the generalization of Fick’s law (cf. Holmqvist et al., in press); (––) that the sensitivity to /vS9O is higher than to /pS9Gout , and the
/pC2 H6 S  5; (–) /pH2 O S. (b) Transient analysis of the integral mean value of absolute effect on /ms S is more pronounced at temperatures
elementary surface reactions /r i S ðmol m2 s1 Þ at Gsub . (–) /r 1 S; (– –) /r2 S. The corresponding to high growth rates.
dotted lines perpendicular to the t-axis represent the individual adsorptive Provided that /pS9Gout , /vS9O , and T are held constant, there
precursors and purge exposures of one pulse sequence. The process operating
are two main ways of varying the half-cycle average exposure
parameters are analogous to those defined in Fig. 5 and T ¼ 1:5  102 1C.
dose for the a th adsorptive precursor (Adomaitis, 2010; Gordon
et al., 2003; Jur and Parsons, 2011):
Z Z
1
on the increment of film mass per unit area, resolved at the level /da S ¼ p @G@t ð12Þ
9Gsub 9 Dta Gsub a
of a single pulse sequence, /ms S was performed. Furthermore, the
model-based analysis is extended to comprehend the mechanistic either by changing the mass flow of the precursors, Q_ a , thereby
parameter impact on the spatially dependent film thickness changing the partial pressure, or by changing the duration of the
profile, conveniently represented by the uniformity factor, UF. pulse, Dt a . Hence, under prevailing operating conditions, the
In Aarik et al. (2006), the influence of carrier gas mass flow, required minimum substrate saturation exposure dose of the
Q_ N2 , was evaluated on the basis of the linear flow rate, and in this a th adsorptive precursor is governed by
study, expressed mathematically as the integral mean value of the Z Z
1 n  ðroa v þ ja Þ
velocity field, /vS9O , in the computational domain O (cf. L¼ @G@t ð13Þ
9Gsub 9 Dtsat, a Gsub Ma
Holmqvist et al., in press):
Z Z As emphasized by Kuse et al. (2003), a net effect of Q_ a and Dt a on
1 1 X 1 /ms S can only be observed provided that the saturation exposure
/vS9O ¼ v@O ¼ r Q_ a @O ð11Þ
9O9 O 9O9 O rAO a ¼ 1 STP, a dose according to Eq. (13) is not achieved. The interdependence of
mass flow and the pulse duration of ZnðC2 H5 Þ2 adsorptive pre-
where AO is the cross section area of O and r ¼ pMðRTÞ1 the local cursor on /ms S is graphically presented in Fig. 9d. It is evident
density. The predicted model response depicted in Fig. 9a shows that /ms S reached a plateau as saturation of the surface sites is
that /ms S decreases, at an applied carrier gas pressure, /pS9Gout , achieved, which demonstrates the characteristic self-limiting
at Gout , when increasing the linear flow rate of the carrier gas. nature of the ALD processes (Lim et al., 2003; Puurunen, 2005).
Analogous dependencies were observed when increasing the As emphasized by Cleveland et al. (2012), the across substrate
carrier gas pressure while maintaining a constant linear flow uniformity is a useful process metric to evaluate the self-limiting
rate, although in this case, the sensitivity of the carrier gas behavior of the ALD process. In agreement with experimental
pressure on /ms S was weaker. This conclusion is consistent with observations in Henn-Lecordier et al. (2011), enhanced uniformity
the experimentally observed parametric behavior in Aarik et al. at higher precursor exposure doses, /dZnðC2 H5 Þ2 S, were obtained in
(2006). Fig. 10d. However, it is noteworthy that the film thickness
326 A. Holmqvist et al. / Chemical Engineering Science 94 (2013) 316–329

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
50 50
180 180
310 310
440 26 15 440 90 50
570 48 37 570 170 130
700 70 59 700 250 210

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
15 1.5
26 1.2
37 0.9
48 90 50 0.6 16 20
59 170 130 0.3 8 12
70 250 210 0 0 4

Fig. 9. Normalized integral mean mass increment /ms S (a.u.) as a function of (a) integral mean value of the velocity field /vS9O ðm s1 Þ on O and pressure at the outlet
/pS9Gout (Pa); (b) process temperature T (1C) and outlet pressure /pS9Gout (Pa); (c) process temperature T (1C) and integral mean value of the velocity field /vS9O ðm s1 Þ on
O; (d) inlet mass flow Q_ ZnðC2 H5 Þ2 (sccm) and pulse duration DtZnðC2 H5 Þ2 (s). The values of the parameters not varied in each case were assigned according to the reference
conditions: temperature T ¼ 1:5  102 1C, pressure at the outlet /pS9Gout ¼ 3:0  102 Pa, integral mean value of the velocity field /vS9O ¼ 3:0  101 m s1 on O, inlet
adsorptive precursor mass flow ½Q_ ZnðC2 H5 Þ2 , Q_ H2 O  ¼ ½5:0,5:0 sccm, pulse sequence ½Dt ZnðC2 H5 Þ2 , Dt N2 , Dt H2 O , Dt N2  ¼ ½2:0,4:0,2:0,4:0  101 s).

uniformity factor, UF, is not strictly increasing with the precursor analyses, design or optimization. For this reason, a systematic
exposure dose. methodology to support the development and statistical valida-
Furthermore, the interdependence of /pS9Gout and /vS9O tion of ALD process models was proposed. Within this framework,
depicted in Fig. 10a reveals that UF approaches unity at the lower the statistical uncertainty of least-squares estimates of kinetic
parameter boundaries. By analysing the spatially dependent film parameters involved in the ALD gas–surface reaction mechanism
thickness profiles depicted in Fig. 3, it is evident that thickness governing ZnO film growth, and its ultimate impact on the model
uniformity enhancement imposes higher reaction driving forces, response have been assessed.
r i pðpca Þ9G and 8i a 3, as z-9Gsub 9 (m). As aforementioned, the The comprehensive model assessment reported here showed
sub
decrease in /pS9Gout and /vS9O promoted enhanced mass flux, ja , that the expected high parameter correlation, originating from
to the substrate surface and mass fraction, oa , of the a th the complex interdependence between elementary surface reac-
precursor, respectively, and thus, explains the parametric depen- tions involved in ALD and the intrinsic mathematical structure of
dence on UF. In addition, the UF temperature dependence resem- the Arrhenius equation, was successfully diminished through
bles that of /ms S (cf. Figs. 9 and 10), however the impact is not reparameterization of the Arrhenius equation, and hence, leading
that pronounced in the lower temperature region, in which the to the improvement in precision of the parameter estimates. It is
overall growth rate is attributed to the activation energy of r i and also shown that the model predictions of the spatially dependent
i A f1; 2g (cf. Section 5.2.1). film thickness profile are in good agreement with both calibration
and validation experimental data, respectively, under a wide
range of operating conditions.
6. Concluding remarks The novel model-based methodology presented in this paper
constitutes a general platform for mechanistic investigations of the
This paper demonstrates the universal applicability of the detailed process dynamics of a continuous cross-flow ALD reactor.
mechanistic model of the continuous cross-flow reactor system The extensive theoretical analysis exhibits that the experimentally
F-120 by ASM Microchemistry Ltd. Suntola (1992) derived in the validated model, successfully reproduces the detailed process
first article of this series (Holmqvist et al., in press). According to dynamics revealed by modern in situ monitoring technologies, as
the ITRS initiative, models of ALD processes must have a high advocated by the ITRS initiative, and thereby provides a supplemen-
level of accuracy if they are to be used to assist in process tary analysis tool. Furthermore, the univariate sensitivity analysis
 A. Holmqvist et al. / Chemical Engineering Science 94 (2013) 316–329 327

a b
1 1

0.96 0.96

0.92 0.92

0.88 0.88

0.84 0.84

0.8 0.8
50 50
180 180
310 310 250
440 59 70 440 210
570 48 570 130 170
26 37 90
700 15 700 50

c d
1 1

0.96 0.96

0.92 0.92

0.88 0.88

0.84 0.84

0.8 0.8
15 1.5
26 1.2
37 0.9
48 210 250 0.6 4 0
59 130 170 0.3 12 8
70 50 90 0 20 16

Fig. 10. Across substrate film thickness uniformity, UF, as a function of (a) integral mean value of the velocity field /vS9O ðm s1 Þ on O and pressure at the outlet /pS9Gout
(Pa); (b) process temperature T (1C) and outlet pressure /pS9Gout (Pa); (c) process temperature T (1C) and integral mean value of the velocity field /vS9O ðm s1 Þ on O;
(d) inlet mass flow Q_ ZnðC2 H5 Þ2 (sccm) and pulse duration Dt ZnðC2 H5 Þ2 (s). The values of the parameters not varied in each case were assigned according to the reference
condition in Fig. 9.

revealed the mechanistic dependence of the measured process g response function –

operating parameters on ALD film growth performance metrics. hs film height m
Hence, the presented model-based framework also serves as a means J parameter Jacobian matrix –
to guide future research efforts in the field of ALD process optimiza- ja diffusive mass flux vector kg m2 s1
tion. On the basis of the characteristic timescales of the mass ki reaction rate constant ðmol m2 Þ1nads,i Pa1 s1 ,
transport and the surface reaction mechanism considered, it is
ðmol m2 Þ1ndes,i s1
possible to determine the optimal process operating parameters to
Ma molar mass kg mol
1
achieve optimal film growth rate and adsorptive precursor utiliza-
tion, while simultaneously controlling film thickness uniformity ms film mass increment kg m2
across the substrate. N~ Avogadro’s number mol
1

ni surface reaction adsorption –

and desorption order
p pressure Pa
Nomenclature
Q_ a volumetric flow rate at STP Nm3 s1
R universal gas constant 1
J mol K1
Roman letters
s2 the variance estimate of s2 –
Ai frequency factor in the ðmol m2 Þ1nads,i Pa1 s1 , T temperature K
Arrhenius equation ðmol m2 Þ1ndes,i s1 t time s
C Nb  N b parameter – t ðj;a=2Þ Student’s t-distribution –
correlation matrix UF spatial film profile –
Dab binary diffusivity m2 s1 uniformity factor
Ei activation energy J mol
1 u design variables –
F system of differential – v velocity vector m s1
algebraic equations W weight matrix –
F ðNb , j;aÞ Fisher’s F-distribution – w algebraic variables –
328 A. Holmqvist et al. / Chemical Engineering Science 94 (2013) 316–329

x state variables – Aarik, J., Siimon, H., 1994. Characterization of adsorption in flow type atomic layer
y measured state variables – epitaxy reactor. Appl. Surf. Sci. 81 (3), 281–287.
Adomaitis, R.A., 2010. Development of a multiscale model for an atomic layer
Greek letters deposition process. J. Cryst. Growth 312 (8), 1449–1452.
Alberton, A.L., Schwaab, M., Biscaia Jr., E.C., Pinto, J.C., 2010. Sequential experi-
mental design based on multiobjective optimization procedures. Chem. Eng.
a significance level –
Sci. 65 (20), 5482–5494.
b calibration parameter – Alberton, A.L., Schwaab, M., Loba~ o, M.W.N., Pinto, J.C., 2011. Experimental design
vector for the joint model discrimination and precise parameter estimation through
g degree of hydroxylation – information measures. Chem. Eng. Sci. 66 (9), 1940–1952.
Asprey, S.P., Macchietto, S., 2000. Statistical tools for optimal dynamic model
G portioned boundary – building. Comput. Chem. Eng. 24 (2–7), 1261–1267.
da half-cycle average Langmuir Bard, Y., 1974. Nonlinear Parameter Estimation. Academic Press, New York.
precursor dose Bates, D.M., Watts, D.G., 1988. Nonlinear Regression Analysis and Its Applications.
John Wiley & Sons, Inc., New York.
dy relative error residual – Beale, E.M.L., 1960. Confidence regions in non-linear estimation. J. R. Stat. Soc. B 22
eb^ ı normalized margin of error – (1), 41–88.
Bird, R.B., Stewart, W.E., Lightfoot, E.N., 1960. Transport Phenomena, 2nd ed. John
z substrate boundary Gsub m
Wiley & Sons, Inc., New York.
local coordinate variable Brereton, R.G., 2003. Chemometrics, Data Analysis for the Laboratory and Chemical
yk fractional surface coverage – Plant, 1st ed. John Wiley & Sons, Inc., Chichester.
of surface species Burton, B.B., Goldstein, D.N., George, S.M., 2009. Atomic layer deposition of MgO
using bis(ethylcyclopentadienyl)magnesium and H2O. J. Phys. Chem. C 113 (5),
L maximum molar mol m2 1939–1946.
concentration of surface Buzzi-Ferraris, G., Manenti, F., 2009. Kinetic models analysis. Chem. Eng. Sci. 64
sites (5), 1061–1074.
Chappell, M.J., Gunn, R.N., 1998. A procedure for generating locally identifiable
t residence time s reparameterisations of unidentifiable non-linear systems by the similarity
n numbers of surface OH – transformation approach. Math. Biosci. 148 (1), 21–41.
groups reacting with each 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
ZnðC2 H5 Þ2 molecule
films using titanium tetraisopropoxide and ozone. J. Vac. Sci. Technol. A 30 (1),
xi surface reaction – 01A150.
stoichiometric coefficient Coello, C.A.C., Lamont, G.B., Veldhuizen, D.A.V., 2007. Evolutionary Algorithms for
r density Solving Multi-Objective Problems, 2nd ed. Springer, New York.
kg m3
Coltrin, M.E., Hsu, J.W.P., Scrymgeour, D.A., Creighton, J.R., Simmons, N.C., Matzke,
R parameter covariance – C.M., 2008. Chemical kinetics and mass transport effects in solution-based
matrix selective-area growth of ZnO nanorods. J. Cryst. Growth 310 (3), 584–593.
F weighted sum of squared – COMSOL AB, 2008. COMSOL Multiphysics User’s Guide, Version 3.5a. COMSOL AB,
Tegnérgatan 23, SE-111 40, Stockholm.
residuals Davis, M.E., 1984. Numerical Methods and Modeling for Chemical Engineers, 1st
j degrees of freedom – ed. John Wiley & Sons, Inc., New York.
ca molar fraction of gaseous – Deminsky, M., Knizhnik, A., Belov, I., Umanskii, S., Rykova, E., Bagatur’yants, A.,
et al., 2004. Mechanism and kinetics of thin zirconium and hafnium oxide film
species
growth in an ALD reactor. Surf. Sci. 549 (1), 67–86.
O spatial domain – Draper, D., 1995. Assessment and propagation of model uncertainty (with
oa mass fraction of gaseous – discussion). J. R. Stat. Soc. B 57 (1), 45–97.
species Draper, N.R., Guttman, I., 1995. Confidence intervals versus regions. J. R. Stat. Soc.
D 44 (3), 399–403.
Elam, J.W., George, S.M., 2003. Growth of ZnO/Al2O3 alloy films using atomic layer
Subscripts and superscripts
deposition techniques. Chem. Mater. 15 (4), 1020–1028.
Elam, J.W., Sechrist, Z.A., George, S.M., 2002. ZnO/Al2O3 nanolaminates fabricated
0 initial value by atomic layer deposition: growth and surface roughness measurements.
a, b gaseous species index Thin Solid Films 414 (1), 43–55.
i surface reaction index Elers, K.E., Blomberg, T., Peussa, M., Aitchison, B., Haukka, S., Marcus, S., 2006. Film
uniformity in atomic layer deposition. Chem. Vapor Deposition 12 (1), 13–24.
ı calibration parameter
Elliott, S.D., 2007. Models for ald and mocvd growth of rare earth oxides. In:
index Fanciulli, M., Scarel, G. (Eds.), Rare Earth Oxide Thin Films. Springer, Berlin,
E calibration and validation Heidelberg, New York, pp. 73–86.
set index Emig, G., Hosten, L.H., 1974. On the reliability of parameter estimates in a set of
simultaneous nonlinear differential equations. Chem. Eng. Sci. 29 (2),
k surface species index 475–483.
ref state at the reference Franceschini, G., Macchietto, S., 2008. Model-based design of experiments for
temperature parameter precision: state of the art. Chem. Eng. Sci. 63 (19), 4846–4872.
Gordon, R.G., Hausmann, D., Kim, E., Shepard, J., 2003. A kinetic model for step
STP standard temperature and
coverage by atomic layer deposition in narrow holes or trenches. Chem. Vapor
pressure Deposition 9 (2), 73–78.
s solid Groner, M.D., Fabreguette, F.H., Elam, J.W., George, S.M., 2004. Low-temperature
Al2O3 atomic layer deposition. Chem. Mater. 16 (4), 639–645.
Hamilton, D.C., Watts, D.G., Bates, D.M., 1982. Accounting for intrinsic nonlinearity
in nonlinear regression parameter inference regions. Ann. Stat. 10 (2),
386–393.
Héberger, K., Kemény, S., Vidóczy, T., 1987. On the errors of Arrhenius parameters
Acknowledgement and estimated rate constant values. Int. J. Chem. Kinet. 19 (3), 171–181.
Henn-Lecordier, L., Anderle, M., Robertson, E., Rubloff, G.W., 2011. Impact of
parasitic reactions on wafer-scale uniformity in water-based and ozone-based
This work has been supported by the Swedish Research
atomic layer deposition. J. Vac. Sci. Technol. A 29 (5), 051509.
Council under Grant no. 2006-3738. Hibbert, D.B., 1993. Genetic algorithms in chemistry. Chemometrics Intelligent
Lab. Syst. 19 (3), 277–293.
Hines, W.W., Montgomery, D.C., Goldsman, D.M., Borror, C.M., 2003. Probability
References
and Statistics in Engineering, 4th ed. John Wiley & Sons, Inc., New York.
Holmqvist, A., Törndahl, T., Stenström, S. A model-based methodology for the
Aarik, J., Aidla, A., Kasikov, A., Mändar, H., Rammula, R., Sammelselg, V., 2006. analysis and design of atomic layer deposition processes—part I: mechanistic
Influence of carrier gas pressure and flow rate on atomic layer deposition of modelling of continuous flow reactors. Chem. Eng. Sci, http://dx.doi.org/10.
HfO2 and ZrO2 thin films. Appl. Surf. Sci. 252 (16), 5723–5734. 1016/j.ces.2012.07.015, in press.
 A. Holmqvist et al. / Chemical Engineering Science 94 (2013) 316–329 329

Jur, J.S., Parsons, G.N., 2011. Atomic layer deposition of Al2O3 and ZnO at atmo- Schwaab, M., Lemos, L.P., Pinto, J.C., 2008b. Optimum reference temperature for
spheric pressure in a flow tube reactor. ACS Appl. Mater. Interfaces 3 (2), reparameterization of the Arrhenius equation. Part 2: problems involving
299–308. multiple reparameterizations. Chem. Eng. Sci. 63 (11), 2895–2906.
Kuse, R., Kundu, M., Yasuda, T., Miyata, N., Toriumi, A., 2003. Effect of precursor Schwaab, M., Pinto, J.C., 2007. Optimum reference temperature for reparameter-
concentration in atomic layer deposition of Al2O3. J. Appl. Phys. 94 (10), ization of the Arrhenius equation. Part 1: problems involving one kinetic
6411–6416. constant. Chem. Eng. Sci. 62 (10), 2750–2764.
Lei, W., Henn-Lecordier, L., Anderle, M., Rubloff, G.W., Barozzi, M., Bersani, M., Schwaab, M., Pinto, J.C., 2008. Optimum reparameterization of power function
2006. Real-time observation and optimization of tungsten atomic layer models. Chem. Eng. Sci. 63 (18), 4631–4635.
deposition process cycle. J. Vac. Sci. Technol. B 24 (2), 780–789. Schwaab, M., Silva, F.M., Queipo, C.A., Barreto Jr., A.G., Nele, M., Pinto, J.C., 2006.
Lim, B.S., Rahtu, A., Gordon, R.G., 2003. Atomic layer deposition of transition A new approach for sequential experimental design for model discrimination.
metals. Nat. Mater. 2 (11), 749–754. Chem. Eng. Sci. 61 (17), 5791–5806.
Lim, J.-W., Park, J.-S., Kang, S.-W., 2000. Kinetic modeling of film growth rates of Semiconductor Industry Association, 2011. Modeling and simulations. In: The
TiN films in atomic layer deposition. J. Appl. Phys. 87 (9), 4632–4634. International Technology Roadmap for Semiconductors, 2011th ed. San Jose,
Matero, R., Rahtu, A., Ritala, M., Leskelä, M., Sajavaara, T., 2000. Effect of water dose CA, pp. 1–45, Avaliable online: /http://public.itrs.net/reports.htmlS.
on the atomic layer deposition rate of oxide thin films. Thin Solid Films 368 Siimon, H., Aarik, J., 1995. Modelling of precursor flow and deposition in atomic
(1), 1–7. layer deposition reactor. J. Phys. IV 05 (C5), 245–252.
McKay, M.D., Beckman, R.J., Conover, W.J., 1979. A comparison of three methods Siimon, H., Aarik, J., 1997. Thickness profiles of thin films caused by secondary
for selecting values of input variables in the analysis of output from a reactions in flow-type atomic layer deposition reactors. J. Phys. D 30 (12),
computer code. Technometrics 21 (2), 239–245. 1725–1728.
Moles, C.G., Mendes, P., Banga, J.R., 2003. Parameter estimation in biochemical Storn, R., Price, K., 1997. Differential evolution—a simple and efficient heuristic for
global optimization over continuous spaces. J. Global Optim. 11 (4), 341–359.
pathways: a comparison of global optimization methods. Genome Res. 13,
Suntola, T., 1992. Atomic layer epitaxy. Thin Solid Films 216 (1), 84–89.
2467–2474.
The MathWorks, Inc, 2010a. MATLAB R2010b Documentation, Version 7.11. The
Nagy, T., Turányi, T., 2011. Uncertainty of Arrhenius parameters. Int. J. Chem.
MathWorks, Inc, 3 Apple Hill Drive, Natick, MA 01760-2098.
Kinet. 43 (7), 359–378. TM
The MathWorks, Inc, 2010b. Optimization Toolbox User’s Guide, Version 7.11.
Park, H.-S., Min, J.-S., Lim, J.-W., Kang, S.-W., 2000. Theoretical evaluation of film
The MathWorks, Inc, 3 Apple Hill Drive, Natick, MA 01760-2098.
growth rate during atomic layer epitaxy. Appl. Surf. Sci. 158, 81–91. TM
The MathWorks, Inc, 2010c. Statistic Toolbox User’s Guide, Version 7.11. The
Pinto, J., Loba~ o, M., Alberton, A., Schwaab, M., Embiruu, M., Vieira De Melo, S.,
MathWorks, Inc, 3 Apple Hill Drive, Natick, MA 01760-2098.
2011. Critical analysis of kinetic modeling procedures. Int. J. Chem. React. Eng.
Törndahl, T., Platzer-Björkman, C., Kessler, J., Edoff, M., 2007. Atomic layer
9 (A87), 1–33.
deposition of Zn1  xMgxO buffer layers for Cu(In,Ga)Se2 solar cells. Prog.
Price, K.V., 1999. An introduction to differential evolution. In: Corne, D., Dorigo, M.,
Photovoltaics: Res. Appl. 15 (3), 225–235.
Glover, F. (Eds.), New Ideas in Optimization. McGraw-Hill, London, pp. 79–108. Vajda, S., Rabitz, H., Walter, E., Lecourtier, Y., 1989. Qualitative and quantitative
Price, K.V., Storn, R.M., Lampinen, J.A., 2005. Differential Evolution: A Practical identifiability analysis of nonlinear chemical kinetic-models. Chem. Eng.
Approach to Global Optimization, 1st ed. Springer, Berlin. Commun. 83 (1), 191–219.
Pritchard, D.J., Bacon, D.W., 1978. Prospects for reducing correlations among Varga, L., Szabó, B., Zsély, I., Zempléni, A., Turányi, T., 2011. Numerical investiga-
parameter estimates in kinetic models. Chem. Eng. Sci. 33 (11), 1539–1543. tion of the uncertainty of Arrhenius parameters. J. Math. Chem. 49,
Puurunen, R.L., 2005. Surface chemistry of atomic layer deposition: a case study 1798–1809.
for the trimethylaluminum/water process. J. Appl. Phys. 97 (12), 121301. Watts, D.G., 1994. Estimating parameters in nonlinear rate equations. Can. J. Chem.
Puurunen, R.L., Vandervorst, W., 2004. Island growth as a growth mode in atomic Eng. 72 (4), 701–710.
layer deposition: a phenomenological model. J. Appl. Phys. 96 (12), Witkowski, W.R., Allen, J.J., 1993. Approximation of parameter uncertainty in
7686–7695. nonlinear optimization-based parameter estimation schemes. AIAA J. 31 (5),
Rahtu, A., Alaranta, T., Ritala, M., 2001. In situ quartz crystal microbalance and 947–950.
quadrupole mass spectrometry studies of atomic layer deposition of alumi- Xu, K., Ye, P.D., 2010. Theoretical study of atomic layer deposition reaction
num oxide from trimethylaluminum and water. Langmuir 17 (21), 6506–6509. mechanism and kinetics for aluminum oxide formation at graphene nanor-
Rahtu, A., Ritala, M., 2002. Reaction mechanism studies on the zirconium chloride– ibbon open edges. J. Phys. Chem. B 114 (23), 10505–10511.
water atomic layer deposition process. J. Mater. Chem. 12 (5), 1484–1489. Ylilammi, M., 1995. Mass transport in atomic layer deposition carrier gas reactors.
Ritala, M., Leskelä, M., 2002. Handbook of Thin Film Materials. vol. 1. Academic J. Electrochem. Soc. 142 (7), 2474–2479.
Press, New York. Yousfi, E.B., Fouache, J., Lincot, D., 2000. Study of atomic layer epitaxy of zinc oxide
Schiesser, W.E., 1991. The Numerical Method of Lines: Integration of Partial by in-situ quartz crystal microgravimetry. Appl. Surf. Sci. 153 (4), 223–234.
Differential Equations, 1st ed. Academic Press, San Diago. Zienkiewicz, O.C., Taylor, R.L., 2000a. The Finite Element Method—Fluid Dynamics,
Schwaab, M., Biscaia, J.E.C., Monteiro, J.L., Pinto, J.C., 2008a. Nonlinear parameter 5th ed., vol. 3. Butterworth-Heinemann, Oxford.
estimation through particle swarm optimization. Chem. Eng. Sci. 63 (6), Zienkiewicz, O.C., Taylor, R.L., 2000b. The Finite Element Method—The Basis, 5th
1542–1552. ed., vol. 1. Butterworth-Heinemann, Oxford.