Académique Documents
Professionnel Documents
Culture Documents
Deposition Reactor
Estudio preliminar en control ptimo multiescalar: Reactor de
deposicin qumica
David Esteban Bernal-Neiraa, Federico Lozano-Santamaraa, Watson Lawrence VargasEscobara & Jorge Mario Gmez-Ramreza
a
Abstract
The dynamic optimization of multiscale systems requires the appropriate and efficient integration between the models of the different
scales, as the description of discrete and continuum phenomena. The optimal control of a chemical vapor deposition (CVD) reactor using
a multiescalar model to describe the phenomena present in the system is studied in this article. The gaseous phase is modelled by
differential balances that make a system of partial differential equations (PDE) and the surface phenomena are modelled using a
kinematic Monte-Carlo (kMC) simulation. The flow of information between the models is bidirectional and is given by the boundary
conditions on the substrate surface. With this representation of the problem the injection velocity and the surface temperature as time
functions are obtained. The results obtained are the ones that allowed the film height with minimal rugosity.
Keywords: Multiscale optimization; optimal control; kinematic Monte-Carlo simulation; chemical vapor deposition reator.
Resumen
La optimizacin dinmica de sistemas multiescalares requiere la integracin apropiada y eficiente entre los modelos de las diferentes
escalas, como la descripcin de fenmenos continuos y discretos. En este trabajo se estudia el control ptimo de un reactor de deposicin
qumica de vapor utilizando un modelo multiescalar para describir los fenmenos presentes en el sistema. La parte continua de la fase
gaseosa se modela por medio de balances diferenciales que forman un sistema de ecuaciones diferenciales parciales; mientras que los
fenmenos de la superficie se modelan por medio de una simulacin cinemtica de Monte-Carlo. El flujo de informacin entre los
modelos es bidireccional y se da en las condiciones de frontera presentes en la superficie del sustrato. Con esta representacin del
problema se obtiene la velocidad de inyeccin de sustrato en funcin del tiempo y la temperatura de superficie, las cuales permiten
alcanzar altura de pelcula deseada con mnima rugosidad.
Palabras clave: Optimizacin multiescala; control ptimo; simulacin cinemtica de Monte-Carlo; reactor de deposicin qumica.
1 Introduction
The transport phenomena have been usually seen from the
macroscale, where the conservation equations are applied over
black boxes, or the mesoscale point of view, that considers the
matter as a continuum and applies the conservation equations over
differential elements of matter making a system of partial
differential equations. In certain cases, these levels of description
may not be enough for certain application in chemical engineering,
which have inherent discrete characteristics as the emulsions,
suspension, porous media, surface phenomena, advanced material
processing and nanotechnology applications. On the other hand,
the transport phenomena in molecular or microscopic scales
consider the discrete characteristics of matter in an atomic,
The way the different scales are coupled is the main limit of
the multiescalar study of different processes because there is not a
systematic and valid way to do it [1, 3]. One way to integrate the
different scales is using reduced order models or models obtained
by data fitting that allow relating the variables from the different
scales [4]. Although these fitting models relate the variables of the
microscale and the mesoscale in the operations ranges where they
were determined, the physical description of the phenomenon and
the capacity of predicting the behavior of the process outside the
fitting range are lost.
This multiscale approximation in order to model the processes
leads to consider the operation, design and control problems using
optimization algorithms. Nevertheless, its application is not direct
because of the difference in the lime and length scales, besides of
the high computational cost associated with the microscale models.
Because of this, the coupling between the different scales must be
done in an efficient way with a bidirectional flow of information
so that the optimization is not limited by the computational
demand of a large-scale problem [5]. The mesoscale processes are
described using partial differential equations (PDE) that after being
discretized using finite differences, finite elements of collocation
methods lead to large scale algebraic problems. These problems
can be solved using deterministic nonlinear gradient-based
optimization algorithms. The microscale description can not be
included directly into the large-scale problems because of its
stochastic nature and the high computational cost that is required
by only one integration step. On the other hand, by using an in-out
model for the microscale, determined outside of the optimization
algorithm, the limitation of the multiscale optimization is
overcome provided that polytopic restrictions are added in the
states to guarantee the validity of the fitted model [6].
The modelling and multiscale dynamic optimization of the
film growth phenomenon in a chemical deposition vapor (CVD)
reactor is studied in this article.
The section 2 presents a theoretical revision about the present
phenomena in the CVD reactor and about multiescalar
optimization. The section 3 describes the modelling in the
mesoscale and the microscale of the reactor, the approximation
used to relate these two scales and the study case proposed to
analyze this phenomenon. The section 4 describes the dynamic
optimization problem, the handle of the PDE and the objective
function. The results obtained are discussed in the section 5 and
the last section concludes the study.
2. Theoretical framework
The topic of this article is the optimal control of a CVD
reactor. For instance, in the next section the theoretical framework
of such kind of reactors and of multiescalar optimization will be
developed. As mentioned before, the phenomenology of this kind
of process requires the integration of phenomena in molecular
scales (that are mainly the surface phenomena in the reactor) and
continuum scales (that include the reactor momentum, mass and
energy transport along the reactor). The first part of this section
will develop the description of CVD reactors and the next part of it
will focus on the multiscale optimization.
2.1 Chemical vapor deposition (CVD) reactors
The chemical vapor deposition (CVD) process is a chemical
process used to produce solid materials of high purity and
performance. This kind of process is very versatile and allows the
production of coating, powders, fibers and monolithic components
of practically any metallic and non-metallic material such as
carbides, nitrides, oxides, carbon and silicon compounds, among
others [7]. In a typical CVD process, a wafer of substrate is
2
Figure 1. Ahrrenius plot for the vapor deposition os silicon using different
precursors.
Source: Pearce, (1983).
temperature
(3)
Continuity equation
1
( ) + ( ) = 0
+
( +
+ )
2
= [
(( )) + 2 ]
( , , , ) = 0; = 0; = {0, , }:
symmetry
= ; { , , }; = :
Side
temperature
= ; { , , }; = :
Upper
(21)
(22)
(23)
(24)
(6)
(7)
(8)
(9)
(20)
(
+
+
)
2
(10)
= [
((
)) + 2 ]
.
The resolution of this set of differential equations is subject to
certain boundary conditions described below.
= 0 ; {0, , }; = : Feed flow.
(11)
(12)
= ; { , , }; = 0: Effluent flow
= 0; { , , }; = : No-slip upper wall
(13)
(19)
(5)
By solving the equations of the continuum phase the values of
temperature, concentration and velocity will be found inside the
reactor. Each one of them will depend on its position and moment
of time. In order to determine the value of these variables the mass,
momentum and energy conservation equations will be solved.
Because of the symmetry of the system, the analysis is reduced
into a two-dimensional analysis, such that the conservation
equations are:
(
+
+
)
2
=
+[
( ( )) + 2 ].
(18)
(4)
(
+
+
)
1
2
( )) + 2 ].
=
+[ (
(14)
Axial
(15)
wall
(16)
wall
(17)
that each site is available for the adsorption, making the adsorption
rate independent of the site.
According to the kinetic theory of gases, for an ideal gas, the
adsorption rate is:
0
=
(25)
22
Where s0 is the sticking coefficient, is the Botzmann
constant, the partial pressure of the precursor, is the total
concentration of sites in the surface, is the molecular mass of
the precursor, is the temperature of the gaseous phase just over
the surface and is the strain rate.
To model the desorption, the local microenvironment of the
molecule and its local activation energy must be considered. Under
the consideration of only first-nearest neighbors interactions, the
desorption rate is given by:
0
(26)
=
exp ( )
2
(27)
0 = 0 exp ( )
(28)
=
exp ( )
2
; = ()
(32)
; = ()
(33)
=0
4
=0
If 0 < <
then a molecule is adsorbed.
(34)
If
+ +
+ +
desorbed.
+
If
<
+
+ +
then a molecule is
(35)
(36)
After determining the controlling step that will happen, in
the case of the adsorption a random place on the lattice is chosen
to determine where the molecule will be absorbed. For the
desorption, a random number will be generated in order to
determine the class of molecule that will be desorbed and then
another random number will determine which of the molecules of
this class will be desorbed. In a similar way, a random number will
determine which molecule will migrate in the case it had to and
then randomly a direction of migration will be determined.
Once the event is determined, the increase in real time is
calculated by the following equation:
ln
(37)
=
+ +
Where is another random generated number between 0 and
1.
Finally, the classes into the lattice must be updated and the
kMC simulation must be continued until reaching a stopping
criteria, which was determines as an amount of events.
The parameters used for the modelling of this system are
reported in the Table 2.
+ +
Table 2.
Parameters used for microscopic simulation [14]
Parameter
Precursor molar mass
Precursor diameter
Strain rate
Sticking coefficient 0
Single bond surface energy
Desorption event-frequency constant 0
Desorption associated energy
Migration associated
Lattice size
Amount
32.06 g/mol
3.7*10-10m
5 1/s
0.1
0.74 eV
1.86*106
0.74 eV
0.4608 eV
50
(38)
(39)
= max (
=
()2
): Rugosity constant. []
(51)
= 0 ; = {0, , }; = {0, , } .
The temperature profile is defined by the gradient of the
temperature between the surface and the feed inlet and
does not have radial effects. The walls of the reactor are
heated in such way that the radial temperature gradients
are avoided.
2
( ) = [
(52)
].
(53)
( ) = 0.
2
(54)
=
>0
1 exp()
(55)
=
+ 2
( = 0) = = 1 + 2 2 = 1 .
(56)
1 exp( )
( = ) = =
+ 2 1
(57)
=
1 exp()
exp (
)
() =
+
1 exp (
)
(58)
1 exp (
)
(59)
0 = [
(( ))].
(60)
( ) = 0.
= 1 ln + 2
(61)
( = ) = 2 = 1 ln .
(62)
( = ) = 2 = 1 ln .
(63)
() =
ln +
.
(64)
ln
ln
2
( + ) 2() + ( )
(66)
2
2
The equality constraints of the optimization problem are the
reduced order models corresponding to the reduced order by
physical simplifications model in the mesoscale and the reduced
order directed by data model in the microscale. The inequality
constraints correspond to the operation conditions shown in the
section 3, which are basically polytopic constraints in the rugosity
and temperature states. The manipulated variables in this problem
(40)
(41)
(42)
= 0
(43)
(44)
(45)
(46)
Where the subscript represents the differential variables
and the subscript the algebraic variables. The objective function
is composed by two terms, a terminal cost ( ) and a mobile
cost, represented in the integral, which takes into account the time
variations.
According to the section 2.2 is not possible to include
directly the microscale model that describes the growth of the
surface by the molecule deposition because of it stochastic nature
and its large computational times. In order to tackle this problem a
reduced order model directed by data is constructed in order to
substitute the original detailed model. The design of experiments
used for the construction of the model consists in a factorial design
where the factors are the temperature and the partial pressure of
the precursor, both limited by the conditions described in the
section 3. The response variables are the change in the surface
height, the growth rate and the surface rugosity. The models used
for the phenomenon description have the form of the eq. (47) and
will be determined by the minimization of the quadratic error.
However, the rugosity depends on the surface configuration,
meaning the amount of particle adsorbed in the last time, making
necessary the use of another approximation as the one in eq. (48)
in order to consider this factor.
= + + +
(47)
(48)
)(, ) (,) + (, )
= (,
Where is any of the following variables: change in the
surface height, growth rate and number of molecules in the lattice.
The description of the phenomena that happen in the
mesoscale is made through a PDE system presented in the section
3. In order to reduce the problem size, the next suppositions are
applied [15].
2
(49)
0 ( ) = [
((
)) + 2 ].
(50)
=
;
are the surface temperature as a function of the radius and the time
and the inlet precursor feed velocity as a function of the time.
The objective control function is formulated as a tracking
function of the average film height compared to the set-point while
penalizing the average rugosity of the surface eq. (67).
2
()]
min = [(() ) +
(67)
Figure 6. Fitting of an in-out model from eq. (47) for the growth rate and
the changes of the surface height. The points represent the obtained data by
the kMC and the lines represent the adjusted model.
Figure 7. Fitting of an in-out model (Ec. 44) for the rugosity coefficients.
The points represent the obtained data by the kMC and the lines represent
the adjusted model.
Figure 10. Rugosity and thin film height profiles for constant precursor
feed
References
[1] W. Vargas, Introduction to Multiscale Transport Phenomena,
Bogot: Universidad de los Andes, 2014.
[2] A. Varshney y A. Armaou, Multiscale optimization using
hybrid PDE/kMC process systems with application to thin
film growth, Chemical Engineering Science, vol. 60, pp.
6780-6794, 2005.
[3] Z. Jaworski y B. Zakrzewska, Towards multiscale modelling
in product engineering, Computers & Chemical
Engineering, vol. 35, pp. 434-445, 2011.
[4] L. Biegler, Y. Lang y W. Lin, Multi-scale optimization for
process systems engineering, Computers & Chemical
Engineering, vol. 60, pp. 17-30, 2014.
[5] M. Salciccioli, M. Stamatakis, S. Caratzoulas y D. Vlachos,
A review of multiscale modelling of metal-catalysed
reactions. Mechanism development for complexity and
emergent behavior, Chemical Engineering Science, vol. 60,
pp. 4319-4355, 2011.
[6] A. Varshney y A. Armaou, Reduced order modelling and
dynamic optimization of multiscale PDE/kMC process
systems, Computers & Chemical Engineering, vol. 32, pp.
2136-2143, 2008.
[7] H. Pierson, Handbook of Chemical Vapor Deposition,
Albuquerque: Noyes Publications, 1992.
[8] M. Costello, D. Tossell, D. Reece, C. Brierley y J. Savage,
Figure 13. Rugosity and thin film height profiles for variable precursor
feed
Notice that the rugosity obtained with this new control scheme
(velocity with variable velocity) is higher to the one obtained with
the constant inlet velocity and the deposition height is less in
almost all the range of times but an uniform surface in the radial
direction is desirable for the final product. This shows that the
operation from the point of view of product quality is more
convenient the results obtained taking the inlet velocity as a
variable.
5 Conclusions and perspectives
This article describes the implementation of an open loop
control to a CVD reactor using dynamic optimization. The model
that describes the operation of the reactor has a multiscale
integration. The description of the adsorption reaction in the
surface is simulated through kMC; and the mass, energy and
momentum transfer in the gaseous phase is described by
differential balances. The integration of the two different scales
9
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
10