Chemical Engineering Science 127 (2015) 374–391

Contents lists available at ScienceDirect

Chemical Engineering Science

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

Reaction factorization for the dynamic analysis of atomic layer

deposition kinetics
Elizabeth M. Remmers, Curtisha D. Travis, Raymond A. Adomaitis n
Department of Chemical & Biomolecular Engineering, Institute for Systems Research, University of Maryland, College Park, MD, United States

H I G H L I G H T S

Model reduction approach for surface reaction dynamics.

Based on a reaction model factorization procedure.

Applicable to atomic layer deposition (ALD) and other thin-film processes.

Provides insight into the structure of differential-algebraic equation models describing thin-film deposition processes.

Presents a new dynamic model for alumina ALD.

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

Article history: We develop a Gauss–Jordan factorization procedure to explicitly separate the slow (deposition), fast
Received 6 August 2014 (equilibrium), and instantaneous (conserved) modes of thin-film deposition models describing the
Received in revised form dynamics of the precursor, surface, and deposition chemical species, focusing primarily on the dynamics
29 December 2014
of atomic layer deposition (ALD) processes. Our reaction factorization procedure provides an unambig-
Accepted 24 January 2015
Available online 4 February 2015
uous means of translating sequences of equilibrium and irreversible reactions characterizing a
deposition system into a low-dimensional DAE system when the reaction kinetics are predicted using
Keywords: transition-state theory. The factorization eliminates redundant dynamic modes; an implicit Euler
Atomic layer deposition procedure then is used to solve the singular-perturbation problem describing the time-evolution of
Transition-state theory
the reaction species on the manifold defined by the combination of the equilibrium relationships and
Model reduction
conserved quantities. An alumina ALD process based on the TMA/water precursor system serves as the
Reaction kinetics modeling
Differential-algebraic systems example used in this work; despite the intense study of this ALD process, several new observations
Perturbation analysis regarding this reaction system are made and a number of new questions are raised.
& 2015 Elsevier Ltd. All rights reserved.

1. Introduction

Atomic layer deposition (ALD) is a thin film deposition process used to create highly conformal films with precise control of thickness
and composition. ALD utilizes a cycle of sequential, self-limiting surface reactions to deposit the desired film one monolayer or sub-
monolayer during each cycle. Typically, a binary sequence of gaseous precursors is used, with purge periods in between to prevent gas
phase reactions. The self-limiting nature of the deposition half-reactions arises from surface saturation due to a finite density of surface
reaction sites, or steric hindrance when precursor ligands remain after the chemisorption reactions.
ALD has a contested history; many trace its origin to Finland the 1970s, when in fact this process had already been developed by
another group in the Soviet Union in the 1960s (Puurunen, 2005). ALD was, and still is, used in the manufacture of electroluminescent flat
panel displays but experienced a resurgence in interest in the 1990s for use in microelectronics processing (Leskelä and Ritala, 2003). The
advantages of ALD over other thin film deposition techniques are precise thickness control, good conformality over high aspect ratio
structures, and low process temperature. These advantages make ALD a desirable process for microelectronics applications including high-
k gate oxide deposition for MOSFETs, DRAM trench capacitor dielectrics, 3D multi-gate field effect transistors, and nonvolatile memory
devices (Leskelä and Ritala, 2003; Kim et al., 2009). Photovoltaic applications include surface passivation layers on crystalline-Si cells

n
Corresponding author.
E-mail address: adomaiti@umd.edu (R.A. Adomaitis).

http://dx.doi.org/10.1016/j.ces.2015.01.051
0009-2509/& 2015 Elsevier Ltd. All rights reserved.
 E.M. Remmers et al. / Chemical Engineering Science 127 (2015) 374–391 375

(Werner et al., 2011), as well as buffer and barrier layers in CIGS cells (Pimenoff, 2012; Holmqvist, 2013). ALD also finds use in numerous
other applications such as nanostructured self-cleaning surfaces (Ng et al., 2008), protective coatings for spacecraft surfaces (Cooper et al.,
2008) and glass displays (Pimenoff, 2012), as well as solid lubricant oxides used in various devices including MEMS (Kim et al., 2009) and
automotive components (Doll et al., 2009).
ALD modeling work falls into two main categories: empirical methods which require fitting parameters based on experiments (such as
Holmqvist et al., 2012, 2013), and first principles (or ab initio) methods which do not. First principles methods allow for the exploration of
novel processes, and can provide information about reaction pathways, substrate effects, and precursor decomposition (Elliott, 2012). First
principles ALD studies frequently utilize density functional theory (DFT) calculations to determine reaction pathways and energetics that
in turn can be incorporated in reaction rate expressions, such as those derived from transition state theory (TST), to determine reaction
kinetics. This kinetics information is used to compare the effect of different surface functional groups on reaction rates (Xu and Musgrave,
2004; Xu and Ye, 2010), or to explain experimentally observed differences in growth rate due to different precursor pulsing sequences in
the deposition of HfO2/Al2O3 mixtures (Nyns et al., 2010). Kinetics information can further be fed into a reactor model or kinetic Monte
Carlo (kMC) simulation to determine the mechanism behind experimentally observed behavior such as the temperature dependency of
growth rate (Deminsky et al., 2004), or to predict desirable operating conditions (Travis and Adomaitis, 2013b).
Despite these extensive efforts into understanding the fundamental aspects of ALD surface reaction mechanisms and integration of
surface reaction models with reactor-scale precursor transport models (see Holmqvist, 2013 for an excellent review), a fundamental
understanding of how one models the surface species dynamics during the different phases of an ALD processing cycle is lacking. Likewise,
an understanding of how these models connect to the more extensively studied chemical vapor deposition (CVD) processes requires
further study. The objective of this paper is to take the first step in decomposing the ALD surface reaction dynamics into those processes
which take place relatively slowly (the deposition modes), those that are fast (the equilibrium reactions), and those processes which are
governed by conservation principles, and so constitute instantaneous processes. Our contribution is the development of a reaction
factorization technique which rigorously determines if the deposition process time scales can be separated and whether the singular
perturbation problem that results from the factorization process results in a well-posed DAE system. We claim that the new approach we
present to modeling ALD kinetics provides a rational path to model development that signals when reaction networks are structurally
incorrect, along with providing other insights into the deposition reaction chemistry.

2. A simple deposition reaction scheme

To begin, consider a simplified deposition reaction where we define a gas-phase monomer M and dimer D precursor species and the
equilibrium relationship between the two. In the context of thin-film deposition processes, each precursor molecule M can be thought of
as containing a single atom of the material (species A) to be deposited. An irreversible monomer reaction with surface site X produces the
deposited film species A:

where g0 and f0 are the net-forward equilibrium and deposition reaction rates, respectively. In our analysis, the rates associated with the
equilibrium reactions are assumed to be much greater than that of deposition and so scaling the time appropriately, g 0 ¼ Oð1=ϵÞ with ϵ{1
and f 0 ¼ Oð1Þ. By-products of the surface reactions are omitted in this simplified analysis. Writing the material balances for each species
give the four ODEs in time:
2 3 2 3
½M 2 1
6 7 " #
dcnc 1 d 6 ½D 7 6 1 0 7
7¼6 7 g0
¼ Snc nr rnr 1 ⟹ 6 6 7 ð1Þ
dt dt 6 7
4 σ ½X 5 4 0  1 5 σf 0
σ ½A 0 1

where σ ¼ deposition surface area=reactant gas volume with units m  1. The matrices c, S, and r contain the species concentrations,
stoichiometric coefficients, and reaction rates respectively. nc and nr refer to the number of components and reactions.
Clearly, two issues arise at this point: (i) we have a higher-dimensional set of ODEs (four) relative to the single rate-limiting deposition
step, and (ii) the monomer concentration dynamic behavior ½MðtÞ is governed by reaction rates of different orders of magnitude. Reducing
(1) to its minimal dynamic dimension is straightforward using the QR factorization or the singular value decomposition. However, because
array S represents reaction stoichiometry, not reaction rates, these decomposition methods will not provide guidance on separating time
scales for the system. Therefore, the approach we pursue is based on the Gauss–Jordan factorization of S (see Vora and Daoutidis, 2001 for
a comparable approach). In (1), both dynamic dimension reduction and timescale separation can be accomplished if there exists a
transformation y¼ Uc in which the objective is to determine a new reactant coordinate system y such that to the greatest extent possible,
each reaction becomes associated with a single new reactant yi A Rnc , i.e.,
" #
Inr nr
Unc nc Snc nr  :
0ðnc  nr Þnr
376 E.M. Remmers et al. / Chemical Engineering Science 127 (2015) 374–391

We deliberately state this as an approximation to allow for off-diagonal elements of the matrix product US that appear in certain cases to
be described later. The case nr 4 nc will not be considered in this work. For our example system, we can factor the ODEs above by hand to
find
2 3 2 3 2 3
½D z0 1 0
6 7 " #
dy d 6  ½M 2½D 7 d 6 7 6 7
6 x0 7 6 0 1 7 g 0
¼ 6 7
dt dt 6 7¼ 6 7¼6 7 ð2Þ
4 σ ½X  ½M  2½D 5 dt 4 w0 5 4 0 0 5 σ f 0
σ ½A þ ½M þ 2½D w1 0 0

resulting in the following temporal modes for the specific case k0 ¼ 1 m3 mol  1 s  1, K0 ¼1 m3 mol  1, and σ ¼ 1 m  1:
dz0 d½D 1
fast ðequilibriumÞ ¼ ¼ ð½M2  ½DÞ ð3Þ
dt dt ϵ
dx0 dð  ½M  2½DÞ
slow ðdepositionÞ ¼ ¼ ½M½X ð4Þ
dt dt

instantaneous ðspecies conservationÞ w0 ¼ ½X  ½M  2½D ð5Þ

w1 ¼ ½A þ ½M þ2½D ð6Þ

or, when the factorization (2) can be completed,

dx dUx c dx
slow ¼ ¼ fðcÞ ¼ fðx; z; wo Þ ð7Þ
dt dt dt

dz dUz c 1 ϵ-0
fast ¼ ¼ gðcÞ ⟹ 0 ¼ gðx; z; wo Þ ð8Þ
dt dt ϵ

instantaneous w o ¼ Uw c ð9Þ

where for ϵ-0 we find the semi-explicit differential-algebraic (DAE) system where x are the differential variables, z the algebraic variables,
and yT ¼ ½xT zT wTo  (Biegler, 2000). We observe the “instantaneous” modes correspond to conserved quantities: constant w1 is simply the
total number of deposition element atoms in our control volume and w0 þ w1 ¼ ½X þ ½A arises from the conservation of surface states –
that for every A created by the deposition reaction, one X is consumed.
Again, we note the strong resemblance of our decomposition approach to other nonlinear model reduction procedures (see e.g., Baldea
et al., 2006; Kumar et al., 1998; Sujit et al., 2011; Vora and Daoutidis, 2001). Likewise, this species conservation is at work in systems such
as the isomerization reactions of Wei and Prater (1962) and numerous follow-up studies (e.g., Lombardo and Hall, 1971) where the
eigenvector associated with a single zero eigenvalue of the reaction rate coefficient array represents the conservation of species A1 þ A2 þ A3
in the three-species equilibrium reaction network of the cited studies.

2.1. Time integration

Other than being stiff for ϵ{1, direct numerical integration of (3)–(6) subject to initial conditions co ¼ cðt ¼ 0Þ is relatively
straightforward: (1) the initial conditions are translated to the new chemical coordinate system yo ¼ Uco to determine the initial values
xo , zo , and wo , of the slow, fast, and instantaneous modes, respectively; (2) the ODEs then are integrated in time in terms of x and z,
computing the reaction rates using current values of the concentration determined from c ¼ U  1 y. We note that U must be invertible
because the forward elimination procedure used to compute it is reversible, and that w ¼ wo for all time. Representative results for (3)–(6)
obtained using a fixed step-size implicit Euler integrator are shown in Fig. 1 for co ¼ ½0:4; 0:6; 1; 0T resulting in xo ¼ 1:6, zo ¼0.6, and
wo ¼ ½  0:6; 1:6T illustrating the rapid contraction from the initial condition co to the neighborhood of the equilibrium manifold Q and the
subsequently slower dynamics which follow.
Multiplying (3) through by ϵ, we observe that our reaction system is a singularly perturbed ODE system for ϵ{1: setting ϵ ¼0 results in
the loss of the fast dynamics in species space associated with relaxation to the manifold defined by the equilibrium conditions
Q ¼ fc : K 0 ½M2  ½D ¼ 0g. Let us consider computing the outer solution only of the singular-perturbation problem (Nayfeh, 1981) subject to
initial conditions co in the following manner:

(i) Project the specified initial conditions co onto the equilibrium manifold Q to find c0 by solving the set of nonlinear/linear algebraic
equations:
2 3 2 3
Ux c 0 xo
6 gðc0 Þ 7 6 0 7
4 5¼4 5
Uw c 0 wo

(ii) Given the species state ci at time ti (where i¼0 for the projected initial state) compute:
xi ¼ Ux ci
 E.M. Remmers et al. / Chemical Engineering Science 127 (2015) 374–391 377

Fig. 1. Two-dimensional dynamics with k0 ¼ K 0 ¼ 1, ϵ¼ 0.1, σ¼ 1, and t A ½0; 2. Specified initial conditions and those projected onto the equilibrium manifold (green
dashed curve Q) are marked with the filled blue square and red circle, respectively. Each point on the blue curve denotes 0.02 time units; black dotted lines denote
½M þ 2½D ¼ constant. (For interpretation of the references to color in this figure caption, the reader is referred to the web version of this paper.)

(iii) Compute the implicit Euler update over time step Δt by driving the residual below to zero using a Newton–Raphson or other
technique suitable for nonlinear algebraic equations:
2 3
ðxi þ 1 xi Þ=Δt fðU  1 yi þ 1 Þ
6 7
6 gðU  1 yi þ 1 Þ 7
4 5
wi þ 1 wo

A demonstration of this method in the solution of (3)–(6) for ϵ ¼0 is shown in Fig. 1 where one can see the initial specifications co , the
projection of these conditions into the equilibrium manifold Q to give the initial conditions c0 , and time integration on the manifold
approaching equilibrium point c1 . Because our simple semi-explicit DAE is index-1 for all initial conditions ½Mo a  1=4, consistency of
the initial conditions is not an issue for this example. However, more detailed deposition surface reaction models may have DAE index
values that depend on initial states of the growth surface that are physically realizable (as opposed to the negative concentration for our
demonstration system), necessitating more sophisticated treatment of the initial conditions (Biegler, 2000; Pantelides et al., 1988).

2.2. Multiple (slow) deposition reactions

We can add the following reaction to describe a second mode of deposition directly from the dimer species:
f1
D þ X-A þ M; f 1 ¼ k1 ½D½X

to find that the Gauss–Jordan factorization cannot be completed; however, the result still accomplishes the dimension reduction and time
scale separation objectives. For k0 ¼ k1 ¼ K 0 ¼ 1 we compute
dx0 dð½M þ2½DÞ
slow ¼ ¼  ½M½X  ½D½X
dt dt
dz0 d½D 1 ϵ-0
fast ¼ ¼ ð½M2  ½DÞ  ½D½X ⟹ ½M2 ½D ¼ 0
dt dt ϵ
instantaneous w0 ¼ ½X þ½A
w1 ¼ ½A þ ½M þ 2½D

which are precisely the same conserved quantities (one reaction site X consumed for each surface A produced, and conservation of the
deposited element found in species A, M, and D), the same equilibrium relationship, and a single dynamic mode describing the
consumption of the total precursor supply by the combination of both surface reactions.

2.3. Non-physical solutions

To illustrate a potential numerical challenge introduced by the equilibrium relationships, consider

g0 d½A d½B d½C
A þ B⇌C with ¼ g 0 ; ¼ f 0  g0 ; ¼ g0
dt dt dt
378 E.M. Remmers et al. / Chemical Engineering Science 127 (2015) 374–391

where the concentration of species B is fed by an outside source with constant rate f0. Rewriting the material balances in matrix form
2 3 2 3
½C  1 0 " #
d6 7 6 7 g0
4 ½B 5 ¼ 4  1 15
dt f0
½A 1 0

The factorization of this set of ODEs is trivial (add the first equation to the remaining two), resulting in

½A þ ½C ¼ w0 the instantaneous mode

½B þ ½C ¼ x0 þ Δt f 0 the slow mode evolving over time intervalΔt

On the equilibrium manifold,

½C  K 0 ½A½B ¼ 0

so x0 þ Δt f 0  ½B K 0 ðw0  x0  Δt f 0 þ ½BÞ½B ¼ 0

Without loss of generality, we set K 0 ¼ 1 and find

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
x 0 þ Δt f 0  1  w 0 7 ð1 þ w0  x0  Δt f 0 Þ2 þ 4ðx0 þ Δt f 0 Þ
½B ¼
2

From the initial state we can determine x0 ¼ Bðt ¼ 0Þ þ Cðt ¼ 0Þ Z 0 and so only the positive-root solution can be valid. While this result
is unambiguous, it is nevertheless important because when K 0 c 1, the implicit numerical integration procedure can overshoot the limit of
½A-0 and converge upon the negative, and physically meaningless solution. Therefore, care must be taken in time-step size control in
cases where the f0 terms are significant and the equilibrium relationship favors the reaction product.

3. Alumina ALD

In a sequence of studies by the authors, a detailed model of alumina ALD using trimethylaluminum (referred to in this paper as TMA,
0
Me3 Al , or precursor A) and water (H2 O0 , precursor B) has been developed (Travis and Adomaitis, 2013a,b,c); note that the prime (0 )
notation will be explained in the following section. The current paper reflects the next step in the continued evolution of this model. While
the overall reaction can be written as

2AlðCH3 Þ3 ðgÞ þ3H2 OðgÞ-Al2 O3 ðsÞ þ 6CH4 ðgÞ

the actual sequence of elementary steps is, of course, much more complicated.
During the TMA exposure, the Al precursor (species A) is reversibly adsorbed onto surface hydroxyl groups HO0 or oxygen bridges. The
surface adduct species can undergo a sequence of reversible and irreversible reactions with other surface hydroxyl groups or bare
aluminum sites, ultimately leaving either bare aluminum sites or a monomethyl aluminum surface species MeAl.
During the water exposure, precursor B adsorbs onto the base surface aluminum and MeAl sites and likewise proceeds through a
sequence of reversible equilibrium reactions and irreversible reactions. The combination of water-exposure reactions ultimately
repopulates the surface with HO sites.
Because of the challenges and limitations associated with in situ characterization of the ALD surface state (Cabrera et al., 2014; Dillon
et al., 1995; Kwon et al., 2009; Levrau et al., 2015), researchers utilize quantum chemical computations to evaluate reaction mechanisms,
bond configurations and energetics as well as to help interpret the features of in situ FTIR spectra. In the sections that follow, we will make
use of the energetics and structure information provided in these studies to pose four sequences of alumina ALD surface reactions.
Methods from statistical mechanics then will be used to compute the partition functions Z i necessary to determine the equilibrium
constants Ki based on the translational, vibrational, rotational, and electronic contributions to Z i .

3.1. Reaction sequence R0: hydroxylated alumina reactions during the TMA exposure

0
The adsorption and subsequent reaction of gas-phase (monomer) TMA (Me3 Al ) onto a surface with sufficient HO0 species can proceed
0
through the following sequence of reactions. The prime notation (O0 and Al ) is used to indicate coordinatively unsaturated species; for
 E.M. Remmers et al. / Chemical Engineering Science 127 (2015) 374–391 379

alumina films these are 2-coordinate O and 3-coordinate Al.

In the first reaction E0, gas-phase TMA adsorbs onto a surface hydroxyl site HO0 forming the surface adduct species we denote Me3AlHO
because of the 4-coordinate Al formed by this adsorption reaction. The oxygen likewise increases its coordinate number to three
producing the HO group. In all of the equilibrium and rate expressions to follow, we denote species surface concentration with ½ and
maximum value by the hat ½^ notation.1 We write the net forward reaction rate r0 (in nm  2 s  1) as
!
P ½S3 P ½S3 1
r 0 ¼ r 0 þ  r 0  ¼ k0 þ A ½HO0  3 k0  ½Me3 AlHO ¼ k0  l0 K 0 A ½HO0  3  ½Me3 AlHO ¼ g 0 ð10Þ
kB T ^
½S kB T ^
½S ϵ0

with ϵ0 ¼ 1=k0  and K 0 ¼ k0 þ =k0  by microscopic reversibility (Pitt et al., 1994). Because we do not have values for the forward k0 þ and
reverse rate constants k0  , we will limit the scope of this study to cases in which ϵi -0 and will examine finite adsorption and desorption
rates in a followup study. With this assumption in mind, we write the equilibrium relationship E0 explicitly, noting K0 has units of volume,
as
½Me3 AlHO
E0 : l0 K 0 ¼ m3 ð11Þ
^ 3
ðP A =kB TÞ½HO0 ð½S=½SÞ
^ 3 term; there must be three
In these rate and equilibrium relationships, the effect of Me group steric hindrance is modeled by the ð½S=½SÞ
open surface sites S adjacent to the surface hydroxyl site for TMA to adsorb. The value of l0 represents the statistical factor (Laidler, 1987)
associated with the symmetry changes of TMA during the adsorption process; for the trigonal planar TMA gas-phase species absorbing on
a HO0 site, l0 ¼ 2. This interpretation of l0 is equivalent to the quotient obtained when the individual partition function rotational
contributions are divided by their respective species symmetry number.
The adsorbed adducts can undergo subsequent reactions. In a manner similar to reaction E0, we write the net-production rate of the
‡
activated complex Me3 AlHO associated with the (1–2) H-transfer reaction in E1 as
1 ‡
r1 ¼ ðl1 K 1 ½Me3 AlHO ½Me3 AlHO Þ
ϵ1
1
¼ g1 ð12Þ
ϵ1
with l1 ¼ 3 because any of the three Me groups of the absorbed TMA can participate in the reaction. The equilibrium condition is found
simply as g 1 ¼ 0.
Finally, the rate of reaction I0 is
kB T ‡
f0 ¼ ½Me3 AlHO  nm  2 s  1
h
This reaction is modeled as irreversible because the gas-phase CH4 (MeH) produced by the reaction generally is pumped away from the
growth surface in an ALD reactor.

1 0
This means ½Me3 Al  denotes the number of TMA precursor molecules consumed per unit surface area, not gas phase concentration.
380 E.M. Remmers et al. / Chemical Engineering Science 127 (2015) 374–391

0
The dimethyl species Me2 Al produced by I0 is considered unstable in the presence of neighboring surface HO0 species (Elliott and
0
Greer, 2004). The reaction between Me2 Al and a neighboring surface HO0 to form an energetically favorable 4-coordinate Al is labeled E2;
it is followed by a (1–2) H-transfer reaction involving equilibrium reaction E3 and ejection of a methane molecule in reaction I1. We write
the net-forward reaction rate r2 for E2 as
!
0
l2 K 2 0 ½HO  1 K2
r2 ¼ ½Me2 Al   ½Me2 AlHO ¼ g 2 ð13Þ
ϵ2 ^
½HO K2 ϵ2
^ in
Note that the equilibrium coefficient appears as 1=K 2 in g2 because K 2 c 1 (see Table 3 in the Appendix) and that the term ½HO0 =½HO
0 ‡
(13) is used to approximate the probability of a HO0 group being located adjacent to the Me2 Al . The activated complex ½Me2 AlHO 
generation rate then is computed from
1 ‡ 1
r3 ¼ ðl3 K 3 ½Me2 AlHO ½Me2 AlHO Þ ¼ g3 ð14Þ
ϵ3 ϵ3
which is then followed by reaction I1
kB T ‡
f1 ¼ ½Me2 AlHO : ð15Þ
h
The E2 þE3 þI1 reaction subsequence produces species MeAlHO that initiates reaction subsequence E4 þE5 þI2; with
!
0
K4 0 ½HO  1
r4 ¼ l ½MeAl   ½MeAlHO
ϵ4 4 ^
½HO K4
K4
¼ g4 ð16Þ
ϵ4
‡
again noting the use of 1=K 4 because K 4 c 1. We compute the activated complex ½MeAlHO  net production rate with
1 ‡ 1
r5 ¼ ðl5 K 5 ½MeAlHO  ½MeAlHO Þ ¼ g5 ð17Þ
ϵ5 ϵ5
to find the rate of I2
kB T ‡
f2 ¼ ½MeAlHO : ð18Þ
h
0 0
Finally we observe that 3 Al surface species are created over this reaction sequence, with generation of the three Al evenly distributed
between the three irreversible reaction steps. This must be the case because the overall reaction sequence effectively “uncovers” two Al
0
species to which the HO0 were bonded, and one new Al is generated by the reaction sequence.

3.2. Reaction sequence R1: the subsequent water reactions

0
The Me2AlHO and MeAlHO surface species ultimately will react and form Al if the TMA exposure and/or purge period following the
exposure of a hydroxylated surface to TMA is sufficiently long. As described in Travis (2014), water readily adsorbs on bare alumina
surfaces, regenerating the surface hydroxyl HO0 groups. For this study, we incorporate reaction R22 and a modification of R23 of her work
(p. 86) to describe water adsorption and subsequent reaction in the following manner:

0
This water sequence begins with the adsorption of H2 O0 onto an Al site in a Lewis acid–Lewis base adduct formation reaction
essentially identical to E0. The AlH2O adduct forms an activated complex by reacting with a neighboring O0 , which we postulate must
0
constitute one bond of a surface Al . The irreversible reaction I3 produces two surface hydroxyl groups; the products of I3 are
0
ð1 þ xÞHO0 þ ð1  xÞHO þ ð1  xÞAl . We note that surface O atoms in I3 are conserved by the addition/subtraction of neighboring surface O
sites; we will later show that the simulations correctly preserve the true Al/O ratio of the deposited alumina film. Additionally, follow-up
 E.M. Remmers et al. / Chemical Engineering Science 127 (2015) 374–391 381

studies will more carefully investigate HO versus HO0 reactions. Until then, we take x¼ 1 and find the following rate expressions:
 
1 PB 0 1
r6 ¼ l6 K 6 ½Al   ½AlH 2 O ¼ g 6 ð19Þ
ϵ6 kB T ϵ6
!
0
1 ½Al  1
r7 ¼ l7 K 7 ½AlH 2 O  ½Al2 H 2 O‡  ¼ g7 ð20Þ
ϵ7 ^
½Al ϵ7
with K6 having units of volume, and the rate of I3:
kB T
f3 ¼ ½Al2 H 2 O‡  ð21Þ
h
Note that no steric hindrance effects are considered for the water reactions; surface saturation consists of exhausting all possible water
reactions.

3.3. Reaction sequence R2: TMA with bare alumina

A second route to Al deposition during the TMA exposure is made possible by reaction sequence R2 below, where TMA can adsorb on
bare alumina (Elliott and Greer, 2004) to produce the surface adduct Me3AlO and, ultimately, 3-coordinate O. In this case, the product
species of the initial adsorption reaction is Me3AlO and the steric hindrance effects are identical to the case of TMA adsorption on HO0 :

0
The adsorption site of the first reaction of this sequence is a surface O0 bridge two connecting 3-coordinate surface Al . The resulting net-
forward rate expression
!
0 3
1 P 0 ½Al  ½S 1
r8 ¼ l8 K 8 A ½Al   ½Me AlO ¼ g8 ð22Þ
ϵ8 kB T ^ ½S
½Al ^ 3 3
ϵ8
0
is written in terms of ½Al  because just as in the water reaction discussed earlier, while the TMA reacts with the lone pair of electrons of the
0
surface O0 the O0 must be adjacent to a surface Al site. As with K0, the equilibrium constant K8 has units of volume.
0
The product of reaction E8 proceeds through a barrierless dissociation step E9 to produce dimethyl (Me2 Al O) and monomethyl (MeAl)
surface species. We write the rate of this decomposition reaction as
 
K9 ^  1 ½Me Al0 O½MeAl ¼ K 9 g
r9 ¼ l9 ½Me3 AlO½Al ð23Þ
ϵ9 K9
2
ϵ9 9
0 0
On locally bare growth surfaces (those with low HO0 concentration), Me2 Al O may react dissociatively with an adjacent Al surface site to
0‡
produce two monomethyl aluminum surface species by the E10 þI4 reaction sequence. The activated complex Me2 Al2 surface
concentration and the rate of the subsequent irreversible reaction are computed by
!
0
1 0 ½Al  0‡ 1
r 10 ¼ l K ½Me Al O  ½Me2 Al2  ¼ g 10 ð24Þ
ϵ10 10 10 2 ^
½Al ϵ10

and
kB T 0‡
f4 ¼ ½Me2 Al2 : ð25Þ
h
0
Again, the ratio ½Al =½Al^ is used to represent the fraction of neighboring surface Al sites that are bare. We note that I is considered
4
irreversible due to the large activation energy of the corresponding reverse reaction.

3.4. Reaction sequence R3: water with surface Me

0 0
Widjaja and Musgrave (2002) studied alumina ALD water-exposure reactions where H2 O0 initially adsorbed on Me2 Al and MeHOAl .
However, because the first of these surface species will be found in only minute quantities on the surface and because we do not consider
382 E.M. Remmers et al. / Chemical Engineering Science 127 (2015) 374–391

the second species, our water-exposure mechanism will be based on the more closely related work of Delabie et al. (2012). In the cited
reference, the authors examine water adsorption and (1–2) and (1–4) H-transfer reactions associated with monomethyl surface species
during the water exposure. Our interpretation of the reaction sequence is as follows:

The reaction sequence begins with a water molecule adsorbing onto the Al atom in one of two adjacent MeAl surface species; we note that
the Al of MeAlH2O retains its 4 coordination by breaking one of its O bonds. The adsorbed H2O proceeds to react with the adjacent MeAl
through a (1–4) H-transfer mechanism first by creating the critical complex Me2 Al2 H2 O‡ , and then ejecting MeH in the irreversible I5
0
reaction. In this final reaction step, we characterize the surface Al species created as Al . The MeAlHO produced by I5 ultimately results in
0
the formation of MeH and Al by reactions E5 and I2.
For these reactions, we determine the following rate expressions:
!
1 PB ½MeAl 1
r 11 ¼ l K ½MeAl  ½MeAlH 2 O ¼ g 11 ð26Þ
ϵ11 11 11 kB T ^
½Al ϵ 11

!
1 ½MeAl 1
r 12 ¼ l12 K 12 ½MeAlH 2 O  ½Me2 Al2 H 2 O‡  ¼ g 12 ð27Þ
ϵ12 ^
½Al ϵ12
with K11 having units of volume, and the rate of I5:
kB T
f5 ¼ ½Me2 Al2 H 2 O‡ : ð28Þ
h

3.5. Computing the equilibrium coefficients

For the numerical solutions that follow, equilibrium coefficients Ki for use in the gi expressions defined in Sections 3.1–3.4 for reactions
E0 through E12 must be computed. The equilibrium coefficients can be calculated using statistical mechanics methods and the relative
contributions of entropy and enthalpy, which correspond to the molecular partition functions and reaction energetics, respectively.
For reaction i involving species j with stoichiometric coefficients νj and enthalpy of reaction ΔE0;i , we can write the equilibrium
constant as
ν
K i ¼ e  ΔE0;i =kB T ∏Z j j ð29Þ
0
For example, in equilibrium reaction E0 where a TMA molecule adsorbs onto an HO site to form Me3AlHO, this expression becomes
Z Me3 AlHO
K 0 ¼ e  ΔE0;0 =kB T ð30Þ
Z Me3 Al0 Z HO0

Details regarding the definition and computation of the partition functions are contained in the Appendix.

3.6. The initial alumina surface state

As pointed out in Elliott and Greer (2004), the amorphous alumina film and its surface are challenging to characterize, and the difficulty
is compounded when the surface state must be parameterized by the degree of hydroxylation.
Insight into the surface structure of amorphous alumina of differing degrees of hydroxylation is provided by Adiga et al. (2007). The
cited study corroborates that the majority of Al in the film is 4 coordinate; while 3 coordinate O dominates in the bulk amorphous film, the
ratio of 2:3 coordinate O switches to favor 2 coordinate at the surface. However, both forms are found at the film surface as well as a
combination of 1 and 2 coordinate HO, the first of which we denoted as HO0 , and the second (HO) we reserve for future study. As an aside,
the cited study also demonstrates that the H is generally concentrated in the top 5 Å of the film.
Given these studies of amorphous alumina structure, we define a bare alumina surface to consist of [3]Al and [2]O with surface
0 ^ If we define η A ½0; 1 as the degree of hydroxylation where η ¼1 indicates a fully hydroxylated surface, we will
concentration of ½Al  ¼ ½Al.
consider the initial state of the alumina growth surface to have zero concentration of all surface species except
0 ^
½Al  ¼ ð1  ηÞ½Al
 E.M. Remmers et al. / Chemical Engineering Science 127 (2015) 374–391 383

^
½HO0  ¼ η½Al

^ ¼ 11:5 nm  2 ; we observe this definition leads to ½HO

with ½Al ^ Surface methyl groups will be limited to the close-packing limit of
^ ¼ ½Al.
^ ¼ 7:2 nm  2 .
½S

4. Simulation methods and results

Having defined the alumina ALD process reaction rate expressions, equilibrium constants, and growth surface initial state, we can now
assemble the simulator for a given ALD process recipe. For that final requirement, we will consider process cycles consisting of a 1 s TMA
exposure at 2 Pa partial pressure, 5 s of post-TMA purge, 1 s water exposure at 2 Pa partial pressure, and finally a 5 s post-water purge.
T ¼450 K for this entire study.
The surface, precursor, and product species (with the exception of S, the open-site species) of the reaction sequences R0 to R3 are listed in
Table 2 of the Appendix; the total number of species including S is 22 for this study. As with (1), we write 22 ODEs in time representing the
species balance for the subset of reactions we choose to include to form stoichiometry array S. Our simulation approach proceeds as follows:

1. We factor the system with the forward elimination procedure used in Gaussian elimination, using pivoting only in cases where a zero
coefficient is located on the diagonal, and performing the elimination procedure on all rows below and above the pivot element, if
possible.
2. If a diagonal array of rank equal to the number of reactions is produced, we can then proceed to separate the fast and slow timescales,
formulating a singular perturbation problem by setting ϵi -0; for our alumina ALD process, we assume for now that the equilibrium
reactions are the fast modes and so set ϵi -0, i ¼ 0; …; 12.
3. In situations where off-diagonal elements remain after the factorization, we may still rigorously proceed with solving the singular
perturbation problem provided that no more than one fast mode remains associated with the new species zi and only slow reactions
are found in the balances for the xi.
4. The “specified” initial condition (Co of Fig. 1), corresponding to the surface state described in Section 3.6 and the gas-phase composition
defined by the process recipe are projected onto the equilibrium manifold, satisfying the equilibrium conditions g¼ 0, x ¼ xo , and the
instantaneous (conserved) modes (producing C0 of Fig. 1).
5. The integration of the DAE system can now proceed using a standard DAE solver; for this study we used the relatively simple implicit Euler
procedure described in Section 2.1 so that the tendency for the integrator to jump to the infeasible solutions described in Section 2.3 could be
studied and avoided; step-size was set to maintain a maximum solution arc-length increment.
6. The previous two steps are repeated for each exposure and purge period for the desired number of cycles.

4.1. Cycle R0 þ R1

We first examine the case where TMA adsorbs and reacts with surface HO0 sites. The R0 reaction sequence proceeds through a partial
0
degree of completion during the TMA exposure and following purge period; the resulting bare Al sites then react with the gas-phase
water precursor in reaction sequence R1.
The factorization procedure for this reaction cycle results in nf ¼8 fast (algebraic equilibrium), ns ¼4 slow ODEs in time, and ni ¼ 10
instantaneous (conserved quantities) modes for this reaction cycle with nc ¼22 chemical species. For this system, the number of original
reactions is nr ¼ 12 ¼ nf þ ns and thus the factorization procedure results in no off-diagonal elements in US. This case can be seen in Fig. 2
where the structure of the factored stoichiometry array is depicted.

Fig. 2. Representations of the factored stoichiometry arrays US for reaction sequences R0 þR1 (left), R2 þ R3 (center), and R0 þR1 þR2 þR3 (right). The black squares indicate
nonzero elements.
384 E.M. Remmers et al. / Chemical Engineering Science 127 (2015) 374–391

0
As can be seen by the square markers in Fig. 3, the initial state of the surface corresponds to 50% Al , 50% HO0 coverage (η ¼ 0:5) and so
both correspond to surface concentrations ½Al=2 ^ ¼ 5:75 nm  2 . A fraction of the surface HO0 species is immediately consumed by the
projection of the specified initial condition onto the equilibrium manifold and then quite rapidly afterwards by the surface reactions
0 0
producing Me2 Al and Me2AlHO; there is a corresponding increase in Al concentration as these sites are “uncovered” by the TMA reactions
with surface HO0 . As the TMA exposure proceeds, we observe the slower decomposition of Me2AlHO by the corresponding rise in MeAlHO
0
and then Al ; because these slower reactions occur after the first irreversible reaction I0, the overall degree of completion of the R0
sequence depends on the length of the purge period for finite TMA exposures. The dashed green curve corresponds to [S] and illustrates
how the surface reaches a high degree of Me saturation (½S-small) during the TMA exposure, but subsequently rises during the purge
period as the surface reactions continue. It is interesting to consider these relatively slow ALD surface reactions and how their existence
suggests an explanation for experimental observations of purge-time dependent ALD growth, such as how the crystallinity of ALD ZnO
films depends on the length of the purge periods (Miikkulainen et al., 2013).
The subsequent water reactions (starting at t ¼6 s) are extremely aggressive, leading to a near-instantaneous jump in HO0 as a result of
0
gas-phase water reacting with Al sites. Surface Me is consumed resulting in the increase in [S] and the sudden increase in ½HO0  results in
0
the rapid shift of Me2 Al to Me2AlHO which then proceeds through the remaining reactions of sequence R1. The surface begins to
0
repopulate with Al during the post-water purge and essentially all the Me2AlHO is consumed in the purge period.
After 10 repeated cycles, the limit cycle solution shown in Fig. 3, right is found. One can observe that all states at the start and end of the
complete cycle match, indicating steady cyclic operation. Many of the qualitative characteristics of the first cycle of operation remain, with some
0
modulation in the peak Al concentration and a significant increase in the mean surface [MeAlHO]. The limit cycle solution is asymptotically stable
because different initial conditions produce this particular solution. By inspection, because of the significant amount of MeAlHO at the end of each
purge period and because [S] never reaches zero, we expect this solution to correspond to non-saturating ALD growth conditions.

4.2. Cycle R2 þR3

0
We now consider the case where TMA adsorbs and decomposes on bare alumina surface Al sites. The R2 reaction sequence proceeds to
ultimately produce surface MeAl which then reacts with the gas-phase water precursor in reaction sequence R3 to regenerate the bare
alumina surface.
The reaction factorization procedure generates nf ¼6 fast modes, ns ¼3 slow ODEs in time, and ni ¼ 13 instantaneous modes. There are
nr ¼ 9 ¼ nf þ ns original slow and equilibrium reactions, and so as with the previous case, the factorization proceeds to completion leaving
no off-diagonal elements in US (Fig. 2, center).
The first- and limit-cycle solutions for this reaction cycle are shown in Fig. 4. A significant difference between the TMA exposure of this case
0
and the previous is the speed of the reactions: reaction sequence R2 is so fast that the conversion of Al to MeAl is nearly complete during the TMA
exposure and stoichiometrically 1-to-1; [S] also follows the rapid dynamics of this reaction sequence, dropping as TMA adsorbs and reacts. It is
interesting to note that in contrast to the previous case, the initial adsorption and subsequent equilibrium reactions are so aggressive that the user-
set 50/50 initial conditions (marked by the squares) and the corresponding initial conditions on the equilibrium manifold are significantly different
in this reaction sequence.
0
The water reactions are comparably fast, with MeAl consumed by the R3 reaction sequence to return Al and to produce MeAlHO;
0
because the latter decomposes according to the slower E5 þ I2 reaction sequence, [S] and ½Al  continue to slowly increase during the post-
water purge period as MeAlHO decomposes resulting in a decrease in [MeAlHO].
As with the R0 þR1 cycle, the limit-cycle solution shares many qualitative aspects with the first cycle. However, the overall MeAl
concentration during the TMA/purge periods is lowered, and more significantly, the decomposition of MeAlHO continues into the TMA
0
and the post-TMA purge periods, resulting in an increase in ½Al  during the latter period. We also note that, as expected, ½HO0  must remain
constant for all time.

Fig. 3. First cycle (left) and limit-cycle (right) solutions corresponding to η ¼ 0:5 and considering reaction sequences R0 þR1 only. Red zones correspond to TMA exposure,
blue to water exposure, and white to purge periods. (For interpretation of the references to color in this figure caption, the reader is referred to the web version of this paper.)
 E.M. Remmers et al. / Chemical Engineering Science 127 (2015) 374–391 385

4.3. The complete deposition scheme

When both reaction cycles are combined in R0 þR1 þR2 þR3, it is interesting to consider the effects of η on the first-cycle behavior and
explicitly demonstrate that both lead to the same limit-cycle solution; representative results for η ¼ 0.95, 0.05 are shown in Fig. 5.

Fig. 4. First cycle (left) and limit-cycle (right) solutions corresponding to η ¼ 0:5 and considering reaction sequences R2 þ R3 only. Red zones correspond to TMA exposure,
blue to water exposure, and white to purge periods. (For interpretation of the references to color in this figure caption, the reader is referred to the web version of this paper.)

Fig. 5. First cycle corresponding to a nearly fully hydroxylated initial surface (η ¼ 0.95, top left) and nearly bare initial surface (η ¼0.05, bottom left). Red zones correspond to
TMA exposure, blue to water exposure, and white to purge periods. Both initial conditions converge to the same limit-cycle solution (right). (For interpretation of the
references to color in this figure caption, the reader is referred to the web version of this paper.)
386 E.M. Remmers et al. / Chemical Engineering Science 127 (2015) 374–391

Fig. 6. The total of each precursor (TMA and water) consumed and methane produced per unit area for the two initial cycles of the full reaction system (left); closeup for
TMA (right) demonstrating more clearly the specified initial conditions marked by squares, and the corresponding conditions projected onto the equilibrium manifold
(circles). Red zones correspond to TMA exposure, blue to water exposure, and white to purge periods. (For interpretation of the references to color in this figure caption, the
reader is referred to the web version of this paper.)

Table 1
Values of limit-cycle gpc and Al/O film ratio for the three reaction networks studied.

Case gpc (Å/cycle) Al/O ratio

R0 þR1 0.34 0.667

R2 þR3 0.14 0.667
R0 þR1 þR2 þR3 0.26 0.667

The reaction factorization procedure results in nf ¼13 fast (algebraic equilibrium), ns ¼5 slow ODEs in time, and ni ¼4 instantaneous
(conserved quantities) modes for the nc ¼22 chemical species. For the full reaction system, the situation becomes somewhat like that
described in Section 2.2 where off-diagonal elements in US remain (Fig. 2, right); for this case nr ¼ 19 4 nf þ ns which indicates some
interaction between the reactions and chemical species when reaction cycle R0 þR1 is combined with R2 þ R3 in this model.
For the first cycle with η ¼0.95, we observe the expected rapid drop in ½HO0 , but now, instead of the corresponding increase in ½Al  seen
0
0
in sequence R0 þ R1, Al sites are open to TMA adsorption via R2 with subsequent, nearly instantaneous, conversion to MeAl. Because of the
0
relative abundance of HO0 , the surface concentration of Me2 Al remains low as Me2AlHO is produced which subsequently decomposes to
0
MeAlHO. Al concentration increases after the start of the post-TMA purge period as MeAlHO continues its slow decomposition.
As expected, MeAl is almost immediately consumed at the start of the water exposure; the surface is quickly repopulated with HO0
while MeAlHO again continues to decompose into the post-water purge period. After 10 repeated cycles, the limit-cycle solution shown in
Fig. 5 right bears significant resemblance to the dynamic features observed in the η ¼ 0.95 initial cycle.
The nearly bare alumina surface η ¼0.05 initial cycle demonstrates much of the same behavior observed in the R2 þR3 sequence, with
0
the rapid generation of MeAl during the TMA exposure. However, because ½S-0, a finite concentration of Al sites remain during the TMA
exposure. Also, ½HO0  does not remain constant over this cycle. Of course, despite the significant differences in the first cycle behavior, the
limit-cycle corresponding to initial conditions η ¼0.05 is the same as that for η ¼0.95.

4.4. Growth per cycle

Using the notation of Travis and Adomaitis (2013a), growth-per-cycle (gpc) denotes the film-thickness gain during each processing
cycle. If the full ALD cycle is defined in the interval t A ½0; t f Þ and the total TMA precursor consumed per nm  2 of film area is computed
0 0
from the concentrations ½Me3 Al 0 and ½Me3 Al tf at the start and end of each cycle, respectively,
0 0
½Me3 Al tf  ½Me3 Al 0
gpc ¼ 10
^
Δz A\  t\pt˚=cycle
½Al
where Δz ¼ 0:295 nm and corresponds to the thickness of one true monolayer of alumina (Travis and Adomaitis, 2013a).
The total of each precursor consumed and methane produced per unit area for the two initial cycles of the full reaction system are
shown in Fig. 6; these data are used in the gpc calculation, as well as for determining the film Al/O ratio2 and the overall material balance
0
for the system, including the by-product methane species. To be precise, the value of ½Me3 Al 0 corresponds to the specified initial
condition value prior to its projection onto the equilibrium manifold – the points marked by squares at the start of each TMA exposure in
0
Fig. 6, corresponding to Co in Fig. 1. ½Me3 Al tf points are marked by squares at the start of the next TMA exposure. Carrying out this
calculation for each of the three cases studied we find the values tabulated in Table 1. In each case, the Al/O ratio accurately represents that
of Al2O3, however, the gpc variations and values relative to what is experimentally observed must now be investigated.

0 0
2
The Al/O ratio is determined by ð½Me3 Al t f  ½Me3 Al 0 Þ=ð½H 2 O0 t f  ½H 2 O0 0 Þ
 E.M. Remmers et al. / Chemical Engineering Science 127 (2015) 374–391 387

4.5. Self-limiting growth

One of the defining characteristics of ALD is self-limiting growth, indicating that there is a maximum amount of material that can
deposit per each half-cycle. Saturation occurs due to the finite number of reactive surface sites or steric hindrance of precursors with larger
ligands, as is the case with the Me ligands of TMA. When precursor exposures are sufficiently long to allow complete saturation of the
surface, the maximum gpc (which we denote GPC) is observed and any further exposure is unnecessary and may be costly in an industrial
setting.
Experimentally measured saturating growth rates for alumina ALD at T¼ 450 K have been reported as 1.1 Å/cycle by Ott et al. (1997)
and 1.25 Å/cycle by Groner et al. (2004); a more extensive discussion of growth-per-cycle can be found in the review by Puurunen (2005).
Given these values, we observe that all three cases of Table 1 correspond to non-GPC growth conditions of a typical TMA/water ALD
process. In particular, we attribute this to the relatively slow reaction rate of MeAlHO through the E5 þI2 reaction sequence, a reaction
pathway deemed energetically unlikely to take place by Delabie et al. (2012). Evidence supporting this conclusion can be found in
Figs. 4 and 5 where high surface concentrations of MeAlHO are visible in the limit-cycle solutions, effectively blocking sites that could
otherwise adsorb the TMA precursor. It also is interesting to observe that the full reaction network results in a gpc value intermediate
between the first two reaction-cycle cases, providing additional evidence of the interaction between the reactions and species of each
reaction cycle.

4.6. An alternate E5

Each of the precursor doses of 2 Torr (2  106 Langmuir) corresponds to an exposure level that should be sufficient for saturating
growth (Travis and Adomaitis, 2013a). Having identified the (1,2) H-transfer reaction of sequence E5 þI2 as a potential rate-limiting step,
we investigate a recent observation of Elliott et al. (2014) that the activation energy of surface reactions can be strongly influenced by the
presence of neighboring surface-adsorbed species. For example, quantum chemical calculations described in the cited work demonstrated
that an activation energy of a surface reaction taking place during the TMA exposure in alumina ALD can be reduced to a third of the
isolated-species value when a neighboring site contained an adsorbed TMA molecule. Shirazi and Elliott (2014) also describe this
“cooperative effect” in the context of enhanced H2O adsorption and subsequent reaction in Hf2O ALD.
To investigate the potential impact this may have on our simulation, we modify the net-forward rate expression (17) to
!
1 alt ½MeAlHO ‡
r alt ¼ l K ½MeAlHO  ½MeAlHO  ð31Þ
5
ϵ5 5 5 ^
½Al

where the effect of a neighboring MeAlHO is modeled by the ½MeAlHO=½Al ^ term and the original activation energy reduced by a factor of
three is used to compute Kalt5 .
As can be seen in Fig. 7, the effect of this change is profound: essentially all of the MeAlHO produced now is consumed allowing for the
0
build-up of MeAl and Al . Furthermore ½HO0  is now much smaller during the TMA and post-TMA purge periods and the surface is rapidly
repopulated with HO0 during the water dose. We note that

gpc ¼1.16 Å/cycle

and that the Al/O ratio remains at 2/3. We also note that the jumps seen at the start/end of the exposure periods are due to the
instantaneous adsorption/desorption events that take place during the switches from purge to precursor exposure and back.
Finally, we examine the behavior of limit-cycle gpc as a function of the TMA and water exposure times in Fig. 8. The gpc of the
simulation depicted in Fig. 7 is located where the TMA and water both are set to a 1 s exposure time. Between these two figures, we
observe the nuanced definition of saturating ALD – that for our nominal process, the surface processes saturate during the water exposure,

Fig. 7. Limit-cycle dynamics for the full reaction network and the alternative E5 kinetics formulation.
388 E.M. Remmers et al. / Chemical Engineering Science 127 (2015) 374–391

Fig. 8. A map of gpc as a function of TMA and water exposure time in Å/cycle; note how the self-limiting behavior results in the saturated ALD plateau. The white marker
represents the nominal operating recipe used in this study.

but some of the slower time-scale reactions of the TMA exposure do not, resulting in a system close to, but now within, the saturating ALD
plateau of this map.

5. Conclusions

In this paper we develop a Gauss–Jordan factorization procedure to explicitly separate the slow (deposition), fast (equilibrium), and
instantaneous (conserved) modes of thin-film deposition models describing the dynamics of the precursor, surface, and deposition
chemical species. In particular, our attention was focused on atomic layer deposition (ALD) processes which are characterized by their
cyclic dynamic behavior and the complexity of defining reactions on the amorphous films generated by many of these processes. This
work was motivated by the increasing complexity of ALD surface reaction models and the need for reliable and accurate numerical
solution procedures for these systems.
The primary benefit of using the reaction factorization procedure developed is that it provides an unambiguous means of translating
the complicated sequences of equilibrium and irreversible reactions characterizing a deposition system into a low-dimensional DAE
system when the reaction kinetics are predicted using transition-state theory. The factorization process eliminates redundant dynamic
modes; an implicit Euler procedure then is used to solve the singular-perturbation problem describing the time-evolution of the reaction
species on the manifold defined by the combination of the equilibrium relationships and conserved quantities.
While it was shown that the physical meaning of the conserved quantities can be clearly discerned in relatively simple deposition
systems, the same cannot be said for more complex systems. Clearly, this is an issue for further study. Likewise, deposition systems with
parallel deposition reactions can result in an incomplete factorization; while this does not affect the validity of the singular perturbation
problem, the physical interpretation of what this means also remains to be investigated.
An alumina ALD process based on the TMA/water precursor system served as the prime example used in this work; despite the intense
study of this ALD process, several new observations were made and a number of new questions were raised. The considerable body of
experimental alumina ALD studies indicates the process settles to a unique and steady limit-cycle solution after a relatively few number of
ALD cycles over most substrates. These observations were found to be consistent with the numerical simulations presented in this study.
Likewise, insufficient exposure to either precursor is known to result in sub-saturating growth, a fact consistent with our simulations. Our
simulations, however, reveal how some reactions can continue into the purge periods following each precursor exposure. This
phenomenon also has been observed experimentally and so our simulations provide new insight into which reactions may be responsible
for this aspect of ALD.
An important, yet unresolved, issue with modeling ALD processes that produce amorphous films is how one represents the growth
surface. We address this problem by formulating the full set of adsorption, desorption, and surface reactions into two reaction mechanism
cycles, each of which can produce an independent limit-cycle solution. These reaction cycles overlap both in terms of the chemical species
that participate in each as well as a subset of the elementary reactions themselves. Combining the individual cycles into one simulation
also produced limit-cycle solutions and revealed the interplay between each reaction cycle. While successful, one of the most important
directions for future study will be how the reaction factorization procedure can be extended to further analyze the structure of ALD
reaction networks to reveal how more complicated mechanisms can be developed and whether valid ALD reaction mechanisms that do
not explicitly enforce this cyclic structure exist.

Acknowledgments

The authors gratefully acknowledge the support of the National Science Foundation through Grants CBET1160132 and CBET1438375.
 E.M. Remmers et al. / Chemical Engineering Science 127 (2015) 374–391 389

Appendix A. Definition and computation of the partition functions

Because experimentally derived equilibrium data are unavailable, we must compute the partition functions for each species involved in
the equilibrium relationships corresponding to (29). We compute the partition functions accounting for the most significant energy
storage modes which are gained or lost during reaction. Modes which are unaffected by reactions are assumed to cancel in the partition
function ratios of the equilibrium relationships, thus they may be omitted in the partition function calculations.
Under the assumption that the translational, rotational, vibrational and electronic states are decoupled, the partition functions Z i are
separable as
Z ¼ ztrans zrot zvib zel ð32Þ
thus each contribution can be computed individually. Our approach follows standard chemical kinetics texts such as Laidler (1987); to
begin, the translational partition function of a polyatomic gas molecule is given by the following equation, which is derived by formulating
the molecule as a particle in a box:
 3=2
2π mkB T
ztrans ¼ 2
ð33Þ
h
where m is the mass of the molecule, and kB and h are the Boltzmann and Planck constants, respectively. The translational partition
function has units of m  3 , while all other partition function components that follow are dimensionless.
For a polyatomic gas molecule with symmetry factor σ and moments of inertia I x ; I y ; I z the rotational partition function is
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pffiffiffiffi  3
π 8π kB T
zrot ¼ Ix Iy Iz ð34Þ
σ h
2

where the symmetry number σ is the number of indistinguishable configurations that can be obtained by rotating the molecule in space.
Alternatively, we can omit the σ term for each species and designate the statistical factor li described in Section 3.1 and in Laidler (1987)
for each equilibrium coefficient. This is, in fact, what we do because of the reduced possibility of inconsistencies that might be created
between different species and our assumptions of what internal rotational modes will cancel.
To calculate the partition function for internal rotations of molecules, hindered rotations are often approximated as harmonic
oscillators or free rotors. The hindered rotation is accurately described by a free rotor at low frequencies and by a harmonic oscillator at
higher frequencies. At low frequencies, the harmonic oscillator approximation overestimates the partition function, approaching infinity
as the frequency goes to zero, while the free rotor approximation overestimates the partition function at higher frequencies (Ayala and
Schlegel, 1998). We expect the internal rotational modes to have relatively low frequencies, so we use the free rotor formulation to
approximate these modes. Thus, the partition function for internal rotational modes depends upon the moment of inertia Ii and symmetry
factor σi of each rotational mode as
pffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2 ffi
π 8π kB T
zrot;internal ¼ ∏ Ii ð35Þ
∏σ i h
2

The vibrational partition function is the product of the contributions from each vibrational mode with frequency νi as the following:
 hνi =2kB T
e
zvib ¼ ∏  hνi =kB T
ð36Þ
i 1e

where the vibrational modes are assumed to be harmonic oscillators.

The electronic contributions are accounted for in the e  ΔE0;i =kB T term of the equilibrium constant, so we do not explicitly calculate this
portion of the partition function.

A.1. Sample computation and resulting K eq values

To illustrate the computation procedure, details follow for the calculation of the partition functions for equilibrium reaction E0, and
these values are listed in Table 2.
The molecular weight of TMA is used in (33) to calculate the translational contribution. The symmetry number σ is set to unity (we will
account for symmetry in the definition of K0) and three moments of inertia I x ¼ I y ¼ 1:468  10  45 m2 kg and I z ¼ 2:936  10  45 m2 kg are
used in (34). These moments are calculated from geometrical considerations (i.e., bond lengths and angles) and the masses of each atom.
We assume that the vibrational modes available in TMA will still be available after adsorption, so they are not calculated explicitly and this
contribution to the partition function is left as unity.
The HO0 adsorption site has no translational freedom and negligible rotational inertia; furthermore, we assume that the O–H bond
vibration remains after adsorption, so its partition function is unity.
Upon adsorption, Me3AlHO has no translational freedom, but two available rotations. The TMA moiety may rotate about the new Al–O
bond with a moment I ¼ 2:609  10  45 m2 kg, or the entire group may rotate about the bond between the surface and the oxygen atom
with a moment of I ¼ 3:603  10  45 m2 kg and a symmetry number σ ¼1. There is a new vibrational mode created by formation of the
Al–O bond, with a frequency estimated to be 1:078  1012 Hz by
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 j Eo j ro
νo ¼ pffiffiffiffiffiffi ð37Þ
2π mads δ3

where Eo is the adsorption energy of the bond under consideration, ro is the bond length, mads is the mass of the vibrating group, and δ3 is
the volume per surface site, calculated from the density and molecular weight of the surface material. Further details on the computation
390 E.M. Remmers et al. / Chemical Engineering Science 127 (2015) 374–391

of (37) can be found in Travis and Adomaitis (2013c). The numerical results for this example and all other species are summarized
in Table 2.
Partition function calculations for all species proceed in this manner, and are used in (29) along with energetics to calculate the
equilibrium constants. These values are listed in Table 3. For each reaction sequence we attempt to obtain all reaction energies from a
single literature source to maintain self-consistent values of equilibrium constants. Note, however, that reaction sequence R0 uses reaction
energetics from multiple sources because some transition-state activation energies were cited as estimates in studies relevant to that
reaction sequence (see Travis and Adomaitis, 2013c for more details).

Table 2
Partition function values for T ¼450 K. All rotational contributions are computed with symmetry factor σ set to unity. Note that data for MeH are not needed because it is
produced by irreversible reactions.

Species ztrans zrot zvib Zi

Me3 Al
0
1.098  1033 m  3 7.591  107 1 8.334  1040 m  3
HO0 1 6.936 1 6.936
Me3AlHO 1 2.767  106 8.691 2.405  107
Me3 AlHO
‡ 1 8.207  103 1 8.207  103
0
Me2 Al 1 4.252  103 5.109 2.172  104

Me2AlHO 1 105.0 1 105.0

‡ 1 13.64 1 13.64
Me2 AlHO
0
MeAl 1 13.64 1 13.64
MeAlHO 1 13.64 1 13.64
‡ 1 1 1 1
MeAlHO

H2 O0 1.372  1032 m  3 160.2 1 2.179  1034 m  3

0
Al 1 1 1 1
AlH2O 1 9.012 3.608 32.51
Al2 H2 O‡ 1 1 1 1
Me3AlO 1 4.362  104 6.636 2.894  105
0 3
Me2 Al O 1 4.252  10 5.109 2.172  104
MeAl 1 13.64 1 13.64
0‡ 1 13.64 1 13.64
Me2 Al2
MeAlH2O 1 69.36 4.202 291.4
Me2 Al2 H2 O‡ 1 13.64 1 13.64

MeH – – – –

Table 3
Equilibrium constant values for T ¼450 K.

Ri Ki Definition Value li ΔE0;i (eV) Source

 26 3
R0 K0 e  ΔE0;0 =kB T l0 Z Me3 AlHO =ðZ Me3 Al0 Z HO0 Þ 5.000  10 m 2  0.7 Elliott and Greer (2004)
R0 K1 e  ΔE0;1 =kB T l1 Z Me
3 AlHO
‡ =Z Me3 AlHO 1:090  10  5 3 0.26 Elliott and Greer (2004)
R0 K2 e  ΔE0;2 =kB T l2 Z Me2 AlHO =ðZ Me2 Al0 Z HO0 Þ 1.133  103 2  0.36n Elliott and Greer (2004)
R0 K3 e  ΔE0;3 =kB T l3 Z Me ‡ =Z Me2 AlHO 1.612  10  13 2 1.09n Elliott and Greer (2004)
2 AlHO

R0 K4 e  ΔE0;4 =kB T l4 Z MeAlHO =ðZ MeAl0 Z HO0 Þ 3.180  1013 2  1.17n Elliott and Greer (2004)
R0 K5 e  ΔE0;5 =kB T l5 Z MeAlHO‡ =Z MeAlHO 2.067  10  15 1 1.21 Delabie et al. (2012)

R1 K6 e  ΔE0;6 =kB T l6 Z AlH2 O =ðZ H2 O0 Z Al0 Þ 3.032  10  22 m3 1  1.01 Hass et al. (1998)
R1 K7 e  ΔE 0;7 =kB T
l7 Z Al2 H2 O‡ =ðZ AlH2 O Z Al0 Þ 5.309  10  3 2 0.095 Hass et al. (1998)

 3
R2 K8 e  ΔE 0;8 =kB T
l8 Z Me3 AlO =ðZ Me3 Al0 Z O0 Þ 1.910  10 m 2  1.2 Elliott and Greer (2004)
R2 K9 e  ΔE0;9 =kB T l9 ðZ Me2 Al0 O Z MeAl Þ=Z Me3 AlO 9.270  104 3  0.4 Elliott and Greer (2004)
R2 K10 e  ΔE0;10 =kB T l10 Z Me 0‡ =ðZ Me2 Al0 O Z Al0 Þ 1.817  10  11 2 0.7n Elliott and Greer (2004)
2 Al2

 25 3
R3 K11 e  ΔE0;11 =kB T l11 Z MeAlH2 O =ðZ H2 O0 Z MeAl Þ 4.387  10 m 2  0.75 Delabie et al. (2012)
R3 K12 e  ΔE0;12 =kB T l12 Z Me2 Al2 H2 O‡ =ðZ MeAlH2 O Z MeAl Þ 1.250  10  6 2 0.33 Delabie et al. (2012)

Values of ΔE0;i that are estimated from their original sources are marked with an n.

References

Adiga, S.P., Zapol, P., Curtiss, L.A., 2007. Structure and morphology of hydroxylated amorphous alumina surfaces. J. Phys. Chem. C 111, 7422–7429.
Ayala, P.Y., Schlegel, H.B., 1998. Identification and treatment of internal rotation in normal mode vibrational analysis. J. Chem. Phys. 108, 2314.
Baldea, M., Daoutidis, P., Kumar, A., 2006. Dynamics and control of integrated networks with purge streams. AIChE J. 52, 1460–1472.
 E.M. Remmers et al. / Chemical Engineering Science 127 (2015) 374–391 391

Biegler, L.T., 2000. Differential-Algebraic Equations (DAEs). Lecture Notes.

Cabrera, W., Halls, M.D., Povey, I.M., Chabal, Y.J., 2014. Surface oxide characterization and interface evolution in ALD of alumina on InP(100) studied by in situ IR spectroscopy.
J. Phys. Chem. C 118, 5862–5871.
Cooper, R., Upadhyaya, H.P., Minton, T.K., Berman, M.R., Du, X., George, S.M., 2008. Protection of polymer from atomic-oxygen erosion using Al2O3 atomic layer deposition
coatings. Thin Solid Films 516, 4036–4039.
Delabie, A., Sioncke, S., Rip, J., Van Elshocht, S., Pourtois, G., Mueller, M., Beckhoff, B., Pierloot, K., 2012. Reaction mechanisms for atomic layer deposition of aluminum oxide
on semiconductor substrates. J. Vac. Sci. Technol. A 30 (1) 01A127:1–10.
Deminsky, M., Knizhnik, A., Belov, I., Umanskii, S., Rykova, E., Bagatur'yants, A., Potapkin, B., Stoker, M., Korkin, A., 2004. Mechanism and kinetics of thin zirconium and
hafnium oxide film growth in an ALD reactor. Surf. Sci. 549, 67–86.
Dillon, A.C., Ott, A.W., Way, J.D., George, S.M., 1995. Surface chemistry of Al2O3 deposition using Al(CH3)3 and H2O in a binary reaction sequence. Surf. Sci. 322, 230–242.
Doll, G.L., Manesh, B.A., Mohseni, H., Scharf, T.W., 2009. Chemical vapor deposition and atomic layer deposition of coatings for mechanical applications. J. Therm. Spray
Technol. 19, 510–516.
Elliott, S.D., Shirazi, M., Klejna, S., 2014. First principles calculations of substrate-specific reactions in ALD, Paper SD-WeM5. In: 61st AVS International Symposium, Baltimore,
MD.
Elliott, S.D., 2012. Atomic-scale simulation of ALD chemistry. Semicond. Sci. Technol. 27, 074008.
Elliott, S.D., Greer, J.C., 2004. Simulating the atomic layer deposition of alumina from first principles. J. Mater. Chem. 14, 3246–3250.
Groner, M.D., Fabreguette, F.H., Elam, J.W., George, S.M., 2004. Low-temperature Al2O3 atomic layer deposition. Chem. Mater. 16, 639–645.
Hass, K.C., Schneider, W.F., Curioni, A., Andreoni, W., 1998. The chemistry of water on alumina surfaces: reaction dynamics from first principles. Science 282, 265.
Holmqvist, A., 2013. Model-Based Analysis and Design of Atomic Layer Deposition Processes (Ph.D. thesis). Lund University.
Holmqvist, A., Törndahl, T., Stenström, S., 2012. A model-based methodology for the analysis and design of atomic layer deposition processes—Part I: mechanistic modelling
of continuous flow reactors. Chem. Eng. Sci. 81, 260–272.
Holmqvist, A., Törndahl, T., Stenström, S., 2013. A model-based methodology for the analysis and design of atomic layer deposition processes—Part II: experimental validation
and mechanistic analysis. Chem. Eng. Sci. 94, 316–329.
Kim, H., Lee, H., Maeng, W.J., 2009. Application of atomic layer deposition to nanofabrication and emerging nanodevices. Thin Solid Films 517, 2563–2580.
Kumar, A., Christofides, P.D., Daoutidis, P., 1998. Singular perturbation modeling of nonlinear processes with nonexplicit time-scale multiplicity. Chem. Eng. Sci. 53,
1491–1504.
Kwon, J., Dai, M., Halls, M.D., Langereis, E., Chabal, Y.J., Gordon, R.G., 2009. In situ IR characterization during ALD of lanthanum oxide. J. Phys. Chem. C 113, 654–660.
Laidler, K.J., 1987. Chemical Kinetics, 3rd ed. Harper & Row, New York.
Leskelä, M., Ritala, M., 2003. Atomic layer deposition chemistry: recent developments and future challenges. Angew. Chem. Int. Ed. 42, 5548–5554.
Levrau, E., Van de Kerckhove, K., Devloo-Casier, K., Sree, S.P., Martens, J.A., Detavernier, C., Denooven, J., 2015. In situ IR spectroscopic investigation of alumina ALD on porous
siliva films: thermal versus plasma-enhanced ALD. J. Phys. Chem. C, in press, http://dx.doi.org/10.1021/jp5088288trebuchet.
Lombardo, E.A., Hall, W.K., 1971. Computerized catalytic kinetics: a useful extension of the method of Wei and Prater. AIChE J. 17, 1229–1233.
Miikkulainen, V., Leskelä, M., Ritala, M., Puurunen, R.L., 2013. Crystallinity of inorganic films grown by atomic layer deposition: overview and general trends. J. Appl. Phys. 113,
021301.
Nayfeh, A.H., 1981. Introduction to Perturbation Techniques. J. Wiley, New York.
Ng, C.J.W., Gao, H., Tan, T.T.Y., 2008. Atomic layer deposition of TiO2 nanostructures for self-cleaning applications. Nanotechnology 18, 445604.
Nyns, L., Delabie, A., Pourtois, G., Van ElshochtmVinckier, S.C., Vinckier, C., De Gendt, S., 2010. Study of the surface reactions in ALD hafnium aluminates. J. Electrochem. Soc.
157, G7–G12.
Ott, A.W., Klaus, J.W., Johnson, J.M., George, S.M., 1997. Al2O3 thin film growth on Si(100) using binary reaction sequence chemistry. Thin Solid Films 292, 135–144.
Pantelides, C.C., 1988. The consistent initialization of differential-algebraic systems. SIAM J. Sci. Stat. Comput. 9, 213–231.
Pimenoff, J., 2012. Atomic layer deposition: excellence in thin film coating. Vak. Forsch. Prax. 24, 6.
Pitt, I.G., Gilbert, R.G., Ryan, K.R., 1994. Application of transition-state theory to gas-surface reactions: barrierless adsorption on clean surfaces. J. Physi. Chem. 98,
13001–13010.
Puurunen, R.L., 2005. Surface chemistry of atomic layer deposition: a case study for the trimethylaluminum/water process. Appl. Phys. Rev. 97 121301:52.
Sujit, S.S., Torres, A.I., Daoutidis, P., 2011. Networks with large solvent recycle: dynamics, hierarchical control, and a biorefinery application. AIChE J. 58, 1764–1777.
Shirazi, M., Elliott, S.D., 2014. Atomistic kinetic Monte Carlo study of atomic layer deposition derived from density functional theory. J. Comput. Chem. 35, 244–259.
Travis, C.D., Adomaitis, R.A., 2013a. Modeling ALD surface reaction and process dynamics using absolute reaction rate theory. Chem. Vapor Depos. 19, 4–14.
Travis, C.D., Adomaitis, R.A., 2013b. Dynamic modeling for the design and cyclic operation of an atomic layer deposition (ALD) reactor. Processes 1, 128–152.
Travis, C.D., Adomaitis, R.A., 2013c. Modeling alumina atomic layer deposition reaction kinetics during the trimethylaluminum exposure. Theor. Chem. Acc. 133, 1414.
Travis, C.D., 2014. Model-Based Analysis of Atomic Layer Deposition Growth Kinetics and Multiscale Process Dynamics (Ph.D. dissertation). University of Maryland.
Vora, N., Daoutidis, P., 2001. Nonlinear model reduction of chemical reaction systems. AIChE J. 47, 2320–2332.
Wei, J., Prater, C.D., 1962. The structure and analysis of complex reaction systems. Adv. Catal. 13, 203–392.
Werner, F., Stals, W., Görtzen, R., Veith, B., Brendel, R., Schmidt, J., 2011. High-rate atomic layer deposition of Al2O3 for the surface passivation of Si solar cells. Energy Procedia
8, 301–306.
Widjaja, Y., Musgrave, C.B., 2002. Quantum chemical study of the mechanism of aluminum oxide atomic layer deposition. Appl. Phys. Lett. 80, 3304–3306.
Xu, K., Ye, P.D., 2010. Theoretical study of atomic layer deposition reaction mechanism and kinetics for aluminum oxide formation at graphene nanoribbon open edges.
J. Phys. Chem. C 114, 10505–10511.
Xu, Y., Musgrave, C.B., 2004. A DFT study of the Al2O3 atomic layer deposition on SAMs: effect of SAM termination. Chem. Mater. 16, 646–653.