Journal of Process Control 19 (2009) 371–379

Contents lists available at ScienceDirect

Journal of Process Control

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

Thermodynamics and chemical systems stability: The CSTR case study revisited
A. Favache, D. Dochain *
CESAME, Universit́e catholique de Louvain, 4-6 Avenue G. Lemaître, 1348 Louvain-la-Neuve, Belgium

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

Article history: This paper is dedicated to the stability analysis of the Continuous Stirred Tank Reactor (CSTR) model by
Received 11 January 2008 considering thermodynamics based arguments. Different Lyapunov function candidates related to the
Received in revised form 30 May 2008 thermodynamics, more specifically based on the entropy, the entropy production and the internal energy,
Accepted 16 July 2008
are considered. These provide new insight and physical interpretation in the stability/instability of the
equilibrium points of the CSTR. This includes in particular extension of the results of Georgakis [C.
Georgakis, On the use of extensive variables in process dynamics and control, Chemical Engineering Sci-
Keywords:
ence 41(6) (1986) 1471–1484] for less restrictive conditions on the system dynamics and thermodynam-
Thermodynamics
Entropy
ics, and invariants sets for the stable equilibrium points of the CSTR by considering an internal energy
Stability based Lyapunov function.
Lyapunov Ó 2008 Elsevier Ltd. All rights reserved.
CSTR

1. Introduction
The analysis and design of control algorithms are largely based
on system theory tools, and in particular on the stability ‘‘à la
Lyapunov” which is indeed intrinsically based and justified on en-
ergy considerations, e.g. [19]. It is therefore natural to consider the
thermodynamics theory for the control design of chemical pro-
cesses. However, if in many situations it is rather easy to describe
the Lyapunov theory in terms of energy for electrical and mechan-
ical systems, this is unfortunately not the case for reaction systems.
The link between thermodynamics and system theory has been an
active research area over the years starting with the seminal works
of Aris and Amundson (see, e.g., [18]) and those of Dammers and
Tels [6], Tarbell [17] or Georgakis [11]. More recently, Alonso, Yds-
tie and coworkers have been quite active in exploring this research
area, resulting in very insightful works on the control design of
process systems (see, e.g., [1–3,20,21]), while Rouchon and Creff
provided important results about the flash dynamics [15].
The objective of this paper is to present several old and new re-
sults that aim at linking the thermodynamics and the system the-
ory concepts via the Continuous Stirred Tank Reactor (CSTR) case
study. The choice of the CSTR is indeed driven by its simplicity
while emphasizing typical important features of reaction system
models: it combines energy and mass balances, it is non-linear
and is possibly characterized by multiple steady states, the system
state variables (concentration, temperature) are non-negative, it is
an open system and the balance equations combine kinetics and
hydrodynamics. The CSTR has also been largely studied in the liter-
ature, yet the way to systematically link thermodynamics and sta-
bility theory even for this apparently simple case study remains an
open question. The main reason is probably mainly due to the com-
plexity of the chemical thermodynamics and the difficulty to link
its concepts with those of system theory. As a matter of example,
the notion of entropy is a priori very attractive to analyze the sta-
bility of a reaction system (and therefore of the CSTR model) but, as
it will be illustrated in the present manuscript, its transcription in
terms of Lyapunov based stability is not obvious. It is probably
worth noting that the negativeness of the entropy production var-
iation can be emphasized with linear phenomenological laws while
no general demonstration is available with non-linear ones (which
corresponds indeed to a large class of chemical systems) (e.g. [7,
pp. 53–56]). Symptomatically, so far the efforts to take advantage
of the positivity of the state variables to consider Lyapunov func-
tions that are not the classical quadratic functions (representative
of the energy of mechanical and electrical systems, but not of reac-
tions systems1) have produced very limited results, the most illus-
trative one being the logarithm based function suggested by
Feinberg [8] for isothermal reactors. Our objective in the present
paper is to clarify as much as possible the reasons that makes
the stability analysis with thermodynamical functions so complex.
With that respect, the paper of Rouchon and Creff [15] dedicated
to the flash case study (i.e. without reaction) provides a good basis
for an appropriate explanation: the main difficulty for such a

* Corresponding author. Tel.: +32 10 472378; fax: +32 10 472180.

1
E-mail addresses: audrey.favache@uclouvain.be (A. Favache), denis.dochain@u- For instance, the thermal energy is proportional to the product of the temperature
clouvain.be (D. Dochain). T and of the specific heat Cp, but not, generally speaking, of the square of T.

0959-1524/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jprocont.2008.07.007
372 A. Favache, D. Dochain / Journal of Process Control 19 (2009) 371–379
thermodynamics based stability analysis arise from the reaction
terms in the balance models.
The paper is organized as follows. We shall first introduce in
Section 2 the modelling assumptions, then derive the dynamical
equations for the CSTR from mass and energy balances, as well
as the entropy and entropy production equations. For the sake of
coherence with thermodynamics, we shall consider a reversible
reaction in the CSTR although the classical CSTR model considers
an irreversible reaction: this point will further discussed and moti-
vated at the end of Section 2. The core of the paper is concerned
with the stability analysis of the CSTR and will be performed by
considering first the linearized model of the CSTR and the first
method of Lyapunov (Section 3), then using the second method
of Lyapunov and the related stability theory results (Section 4).
Several thermodynamics based functions will be considered for
the stability analysis. We shall indeed successively consider in Sec-
tion 4 the entropy, the entropy production, and the thermal energy
as Lyapunov function candidates, and provide in each case a de-
tailed analysis of the obtained stability results.
2. Dynamical model of the CSTR
2.1. Mass and energy balance equations
Let us consider a CSTR in which an exothermic reaction in a li-
quid medium takes place and involve two chemical species A and
B:
A  bB
where b is a stoichiometric coefficient. The reactant and the product
are dissolved in an inert I. The inlet flow only contains A and I, and
the volumetric flowrate is such that the liquid volume in the reactor
is kept constant. The reactor is connected to a jacket in which a
cooling fluid at temperature Tw is circulating. A schematic view of
the reactor is given in Fig. 1.
Fig. 1. Schematic view of the CSTR.
Let us consider the following modelling assumptions.
1. The liquid volume V is independent of the dissolved quantities
of A and B, and only depends on the number of moles of inert nI
in the mixture (nI = CIV) with CI the inert concentration.
2. The inert concentration is constant; the reactor is perfectly
mixed.
3. The temperature of the cooling fluid Tw is constant.
4. The thermodynamical model considered here is that of an ideal
liquid mixture, the heat exchange is proportional to the temper-
ature difference between the cooling fluid and the liquid in the
reactor (Q_ ¼ hðT  T w Þ where Q_ is positive if heat is extracted,
and with h the heat exchanger coefficient), the time evolution
is a quasi-static process, and the molar heat capacity Cv of each
species is constant.
5. The reaction rate r is composed of two reaction rates, one for
the forward reaction rf, the other for the backward reaction rb,
with r = rf  rb. Each reaction rate only depends on the temper-
ature and on the ‘‘reactant” concentration, i.e. A for the forward
reaction and B for the backward reaction. This means in partic-
ular that each individual reaction is non-autocatalytic and with-
out inhibition by the reaction product. For the sake of notation
simplicity, since the volume is constant, the concentrations of A
and B is proportional to the quantity of matter, and therefore
r(T, nA, nB) (=rf(T, nA) rb(T, nB)) can be considered as the reaction
rate per unit volume. Each subreaction rate is an increasing of
the temperature and of the associated quantity of ‘‘reactant”
such that it is equal to zero when the associated ‘‘reactant” is
equal to zero, i.e. rf(T, nA = 0) = 0, rb(T, nB = 0) = 0.
Since the CSTR is a simple quasi-static thermodynamical sys-
tem, its dynamics can be deduced from mass balances on the three
involved species, the internal energy balance and the volume bal-
ance. Since the volume and the inert concentration are constant,
the dynamics of the CSTR are therefore described by the following
three balance equations:
dnA q in
¼ ðC A V  nA Þ  rðT; nA ; nB ÞV ð1Þ
dt V
dnB q
¼  nB þ brðT; nA ; nB ÞV ð2Þ
dt V
dU in q
¼ qCin
v ðT in  T 0 Þ þ qðC A u0A þ C I u0I Þ  U  hðT  T w Þ ð3Þ
dt V
where q, C in in in
A , Tin, Cv (= C A C vA þ C I C vI ), u0j (j = A, I), T0 hold for the vol-
umetric flow rate, the inlet concentration of species A, the inlet tem-
perature, the volumetric heat capacity of the inlet liquid, the
reference molar energy of the species j, the reference state temper-
ature, respectively. Note that strictly speaking from a thermody-
namical viewpoint, inlet and outlet enthalpies should been
considered in the internal energy balance Eq. (3), yet the above
writing with the inlet and outlet internal energies is valid under
the above assumptions of liquid medium and constant volume.
The dynamics of the CSTR are therefore described by the three bal-
ance Eqs. (1)–(3).
They have to be completed by a relation linking T, U, nA and nB.
This is done by considering the thermodynamical model of the
ideal liquid for which the internal energy is given by the following
relationship (see e.g. [16]):
X
U¼ nj ðC vj ðT  T 0 Þ þ u0j Þ ð4Þ
j¼A;B;I
If we note that the global liquid heat capacity Cv is equal to
P
Cv ¼ j¼A;B;I nj C vj , the temperature can be expressed as a function
of the other variables as follows:
 A. Favache, D. Dochain / Journal of Process Control 19 (2009) 371–379 373

P
U þ Cv T 0  i¼A;B;I ðni u0i Þ The entropy production (12) can then be rewritten as follows:
T¼ ð5Þ 0
Cv   1    
T  T  T 
o S o S  A in o S oS  1 _
The dynamical equation of the temperature is easily obtained by rS ¼@    FX þ  rn þ  Q ð17Þ
oX  oX  on  oU X T ext
differentiating Eq. (4): X X in X

dU dT X dni Let us now concentrate more specifically on the CSTR example.

¼ Cv þ ½C vi ðT  T 0 Þ þ u0i 
dt dt i¼A;B;I dt Since the process is quasi-static, the entropy balance of the CSTR
can be obtained by considering the Gibbs’ relation:
By considering Eqs. (1)–(3), the temperature dynamics is then given
by the following equation: dS 1 dU lA dnA lB dnB
¼   ð18Þ
dt T dt T dt T dt
dT
Cv ¼ hðT w  TÞ þ qðC in
A C vA þ C I C vI ÞðT in  TÞ þ rVðDr HÞðTÞ ð6Þ The entropy balance equation is then rewritten by combining (18)
dt
with the balance Eqs. (1)–(3) as follows:
with
dS 1  1 q 1 T  Tw
in
ðDr HÞðTÞ ¼ ½C vA ðT  T 0 Þ þ u0A   ½C vB ðT  T 0 Þ þ u0B  ð7Þ ¼ qCin
v ðT in  T 0 Þ þ q C A u0A þ C I u0I  U h
dt T T V T T
the reaction enthalpy.   1 q 1
þ lA  blB rV  ððC in A V  n A Þ lA  n l
B B Þ ð19Þ
T V T
2.2. Entropy and entropy production The entropy of an ideal liquid mixture is given by the following
relationship:
The Gibbs’ relation [16]:  
T X h n i X
i
1 l P T S ¼ Cv ln R ni ln þ ½ni s0i  ð20Þ
dS ¼ dU  dn þ dV ð8Þ T0 i¼A;B;I
N i¼A;B;I
T T T
P
(with dV = 0 here since V is constant) provides the basis for calculat- where N is the total matter quantity in the liquid: N ¼ i¼A;B;I ni .
ing the time variation of the entropy which can be written as The chemical potential of the species i is given by combining
follows: (18) and (20):
   n 
oS T
dS oT S dX li ¼ T ¼ TC vi ln þ RT ln
i
 Ts0i þ C vi ðT  T 0 Þ þ u0i
¼  ð9Þ oni T0 N
dt oX  dt
X   ð21Þ
oT S oT S oS 
¼  ðFin
X  Fout
X Þ þ  rn þ  Q_ ð10Þ The entropy production can be obtained from the entropy balance
oX  on  oU X
X X (i.e. Eq. (19) combined with (21)) by substracting the net entropy
with FjX ¼ ½FjU ; FjA ; FjB T ðj ¼ in; outÞ the inlet and outlet fluxes of X fluxes coming from the environment:
= [U, nA, nB]T, and rn ¼ ½rnA ; rnB T the rates of consumption/formation
of A and B by the chemical reaction, respectively.  heat transfer:
By considering the inlet and outlet entropy fluxes by convection Tw  T
Fin out Q_ Fheat ¼h ð22Þ
S and FS , and the entropy due the thermal exchange T w , the en-
S
Tw
tropy production is therefore equal to:
!  convection:
dS in Q_ out
"   ! #
rS ¼  FS þ  FS ð11Þ T in C in
dt Tw Fconv ¼ qC in C vA ln  R ln A
þ n A s0A
S A
  T0 C in
A þ CI
oT S in out oT S "   ! #
¼  ðFX  FX Þ þ  rn T in C in q
oX  on  þ qC in C ln  R ln A
þ n s S
X X I vI I 0I 
   T0 C in þ C V
oS  1 _ A I
þ  Q  FS þ Fout
in
ð12Þ
oU X T ext S
Consequently the entropy production is equal to:
From the homogeneity of the reactor, we have ðT  T w Þ2 rV
rS ¼ h þ ðlA  blB Þ þ qðC in A C vA
S¼s Fout
S and X ¼ s Fout
X ð13Þ TT w T
 
with s the reactor residence time.
T in  T T
þ C I C vI Þ þ ln
Since the entropy is an homogeneous function of degree 1, we T T in
" ! #
can write C in n 
in A A
 þ qRC A ln  ln
oT S C in
A þ CI
N
S¼  X ð14Þ " ! #
oX  CI n 
X I
þ qRC I ln  ln ð23Þ
This implies that C in
A þ CI
N

oT S out The different sources of entropy can be identified on the basis of
Fout
S ¼  F ð15Þ (23):
oX  X
X
2
Moreover since there is only one inlet flowrate, the above relation  heat transfer: h ðTT wÞ
TT w
(15) is also true for the inlet entropy flux:  chemical reaction: ðlA  blB Þ rVT h i
  heat convection: qðC in T in T
þ lnðTTin Þ
A C vA þ C I C vI Þ
oT S  
T
 
Fin
S ¼  Fin
X ð16Þ C in
oX   mixing: qRC in  lnðnNA Þ þ qRC I ln C inCþC  lnðnNI Þ
A ln C in þC
A I
X in
A I A I
374 A. Favache, D. Dochain / Journal of Process Control 19 (2009) 371–379

8
The second law of thermodynamics tells us that the entropy >
>  B ¼ bðC in
n 
A V  nA Þ
>
>
production is always non-negative. Since the abovementioned ef- < k ðTÞ þ q n
>
 A  kb ðTÞ½bðC in  b in
f V A V  nA  ¼ qC A
fects are independent from each other, the entropy production ð25Þ
> hðT  T w Þ þ qðC C v ðTÞ þ C I C v ðTÞÞðT  T in Þ
> in
for each of them has to be non-negative. For the heat transfer, this >
> A A I
>
:
implies that h P 0, i.e. that the conductive heat transfer has to go ¼ ðkf ðTÞn A  kb ðTÞ½bðC in
A V  n A b ÞVðDr HÞðTÞ
from the warm source to the cold source. For the heat convection
and for the mixing, it is straightforward to check that the above It is well known (e.g. [4,14]) that the system may have up to three
expressions are non-negative. equilibrium points, as it is illustrated in Fig. 2 for which the follow-
For the entropy production linked to the reaction, there is in- ing parameter values and operating conditions have been
deed an apparent contradiction. (lA  blB) is the affinity of the considered
reaction and r is the reaction rate. On one hand, thermodynamics h ¼ 600 J=ðs KÞ; T w ¼ 323 K; q ¼ 2 l=s; V ¼ 100 l;
considers that all reactions are reversible and the equilibrium is
C in
A ¼ 18 mol=l; T in ¼ 323 K
reached when the affinity is equal to zero. Therefore if the affinity
is positive, the reaction should progress towards a production of B,
i.e. r > 0, and if the affinity is negative, the reaction should progress u0A ¼ 4  104 J=mol; u0B ¼ 2:5  104 J=mol; u0I ¼ 0 J=mol;
towards a production of A, i.e. r < 0. This guarantees the positivity C I ¼ 555; 556 mol=l
of entropy production but excludes the existence of irreversible
reactions. On the other hand, kinetics theory considers the exis- C vA ¼ 10 J=ðmol KÞ; C vB ¼ 5 J=ðmol KÞ; C vI ¼ 8 J=ðmol KÞ
tence of irreversible reactions. Moreover the kinetics models usu-
ally considered in chemical engineering very often do not depend s0A ¼ 100 J=ðmol KÞ; s0B ¼ 150 J=ðmol KÞ;
on the affinity: this is indeed due to the fact that kinetics is largely s0I ¼ 80 J=ðmol KÞ; T 0 ¼ 273 K
based on experiments.
Let us calculate the affinity of the reaction:
b ¼ 2; k0f ¼ 34  108 mol=ðl sÞ; Eaf ¼ 88900 J;
1 5
Aff ¼ lA  blB k0b ¼ 1; 8  10 3 s ; 1
Eab ¼ 10 J
 
T
¼ TðC vA  bC vB Þ ln  Tðs0A  bs0B Þ The dynamics of the linearized tangent model around each equilib-
T0 rium point are readily deduced from the mass and energy balance
!
nA b1 Eqs. (1), (2), (6) and the equilibrium points. The state matrix K of
þ ðDr HÞ þ RT ln N the linearized tangent model can be written as follows:
nbB
0 1
 Vq  RA þ RT CuA RB þ RT CuB  CRT
Note that only one term depends on the composition. At a given B v v v C
B C
temperature, the affinity is positive if: K ¼ B bRA  bRT CuAv  Vq þ bRB  bRT CuB b CRT C ð26Þ
@ v v A
!   h CuA h CuB  Vq  h
Cv
v v
nA b1 T
RT ln N P TðC vA  bC vB Þ ln  T Dr S  ðDr HÞðTÞ
nbB T0 by considering
ð24Þ X
dU ¼ Cv dT þ ui dni ð27Þ
i
with:
with ui the partial molar internal energy:
Dr S ¼ bs0B  s0A

oU 
ui ¼ ð28Þ
the entropy variation of the reaction. If the reaction is highly exo- oni T;V
thermic (large (DrH)) and if the reaction is such that DrS > 0 (more
‘‘disorder”), the affinity remains positive for very low concentration
ratios nnAB . This means that the thermodynamics foresees a reversible 6
x 10
reaction whose equilibrium is largely pushed towards high values 5
of the reaction products. The reaction can then be considered to
be practically completed and the entropy production remains posi-
4
tive up to very high conversions that are never reached in practice.
As mentioned in the introduction, for the sake of coherence
with thermodynamics, we shall consider here a reversible reaction. 3
Energy [J]

Note however that the results presented in the present manuscript

are valid for irreversible reactions which can be considered as a
2
particular case of reversible reactions for which the equilibrium
is displaced to very low values of the reactant concentrations.
1

3. Stability analysis based on the linearized model of the CSTR

0
(Lyapunov’s first method)

Let us now analyze the dynamics of the system around its equi- 1
300 350 400 450 500 550 600 650 700 750 800
libria. Let us assume that the reaction kinetics obeys to the mass
Temperature [K]
action law: rV = kf ðTÞnA  kr ðTÞnbB . The equilibrium values n
A , n
B
and T of the three state variables are solutions of the following Fig. 2. Equilibrium points of the CSTR (straight line: left hand side of the third
expressions: equation of (25); sigmoid curve: right hand side of third equation of (25)).
 A. Favache, D. Dochain / Journal of Process Control 19 (2009) 371–379 375

and with with Mj the molecular weight of the species j, and by stoichiometry,
   MA = bMB. Since nI is constant, z1 is therefore proportional to the dif-
oðrVÞ oðrVÞ oðrVÞ
RA ¼ ; RB ¼ ; RT ¼ ference of mass with respect to the mass at equilibrium.
onA eq onB eq oT eq In the abiabatic case (h = 0), the eigenvalues k2 and k3 and the
Dr HðTÞ h associated variables z2 and z3 are equal to
g¼ ; f¼ ; Cv ¼ Cv jeq
Cv Cv q q
k2 ¼   ðRA  bRB Þ þ gRT ; k3 ¼  ð39Þ
V V
The eigenvalues of the state matrix K are equal to:
~ A þ RB n
z2 ¼ RA n e
~ B þ RT Te ; z3 ¼ U ð40Þ
q
k1 ¼  ð29Þ The result is completely similar to the one in [11], i.e. the three vari-
V
q 1 pffiffiffiffi ables zi (i = 1, 2, 3) are related to the deviations to the total mass, the
k2 ¼  þ ½f  ðRA  bRB Þ þ gRT þ q ð30Þ reaction rate, and to the internal energy, respectively. This means
V 2
q 1 pffiffiffiffi that, in line with [11], we can conclude that the stable modes are
k3 ¼  þ ½f  ðRA  bRB Þ þ gRT  q ð31Þ associated to the total mass and internal energy and the unstable
V 2
is associated to the reaction rate.
with the discrimant q: Note that the result is valid by only assuming that the molar
internal energies only depend on the temperature, an assumption
q ¼ ðf þ ðRA  bRB Þ  gRT Þ2  4fðRA  bRB Þ ð32Þ
used to write that
k1 is obviously negative. It is straightforward to note that k2 is the oT 1 oT ui
first eigenvalue to test to evaluate the (in)stability of each equilib- ¼ ; ¼ ð41Þ
oU Cv oni Cv
rium point. It is then easy to show on the basis of the right hand side
of the temperature Eq. (6) that k2 is negative for the extreme equi- With that respect, the above result extends that of [11] for less
librium points while it is positive for the intermediate one. restrictive assumptions on the system dynamics and
In his paper of 1986 [11], Christos Georgakis has considered thermodynamics.
extensive thermodynamic variables to analyze and give a physical
interpretation of the slow and fast modes of different chemical 4. Stability analysis based on the Lyapunov’s second method
process case studies, and in particular to an isothermal CSTR and
a simplified adiabatic CSTR. The argument is based on the rewriting 4.1. The entropy as the Lyapunov function candidate
of the linearized tangent model by considering the state
transformation: As already mentioned above, the second law of thermodynam-
ics suggests that entropy related functions should be a priori good
z ¼ CT x ð33Þ candidates as Lyapunov functions in order to analyze the stability
with C the eigenvector matrix associated to the state matrix of the properties of the CSTR model. Since the entropy production is as-
linearized tangent KT. Then the dynamics in z are equal to sumed to be positive, one possible option could be to consider
the integral of the trajectories of the entropy production. This func-
dz tion W(U,nA,nB) would be such that
¼ Dz ð34Þ
dt
oT W dX
with D the diagonal matrix of the eigenvalues of K. In his simplified
rS ¼ ð42Þ
oX dt
adiabatic reactor example, Georgakis shows that two entries of the
with X = [U, nA, nB]T. Such a function exists for an isolated system.
vector z are associated to the system reaction invariants (see e.g.
The entropy production is indeed given by
[9,10]) while the third one is associated with the reaction rate
and is a mode of the reaction variant subspace. dS oT S dX
Let us apply this approach to our CSTR example. The variables z
rS ¼ ¼ ð43Þ
dt oX dt
are then written as follows:
since the entropy fluxes are equal to zero. Thermodynamics tells us
~A þ n
z1 ¼ b n ~B ð35Þ that the stationary state of an isolated system is the one for which
the entropy is maximum. Therefore the entropy is an appropriate
q   u u R e

~ B  RT A n
~ A þ RB n ~ A  RT B n
~B þ T U Lyapunov function for isolated systems. However, it is worth noting
z2 ¼ 2Cv þ k2 RA n
V Cv Cv Cv that the entropy production is not the total derivative of a state
 # function for open systems. Indeed the entropy production is then
2h uA u R e
þ RT ~ A  RT B n
n ~B þ T U ð36Þ not equal to the entropy variation due to the presence of the inlet
Cv ðVq þ k2 Þ Cv Cv Cv
and outlet entropy flows
!
q   u u R e

dS Q_
z3 ¼ 2Cv þ k3 ~ A þ RB n
RA n ~ B  RT A n~ A  RT B n
~B þ T U rS ¼  Fin
S þ  Fout
S ð44Þ
V Cv Cv Cv dt T ext
 #
2h u u R e with Fin out
þ q  RT A n ~ A  RT B n ~B þ T U ð37Þ S and FS the inlet and outlet convection flowrates, respec-
Cv V þ k3 Cv Cv Cv tively. The entropy production is not the exact derivative of a state
e are the deviation variables in the linearized tan- function and it is therefore not possible to deduce a Lyapunov func-
where n~A , n
~ B and U
tion for the CSTR model by integration of the entropy production.
gent model of the system of nA, nB and U, respectively. The unstable
mode is associated to z2 while the other two are stable.
4.2. The entropy production as the Lyapunov function candidate
In the above general case, it is possible to give a physical inter-
pretation to z1. It is indeed proportional to the total mass in the
A consequence of the second principle of thermodynamics is
reactor m:
that the entropy is a concave function of the other extensive quan-
A þ n
m ¼ z1 M B þ M I nI þ ðbn B ÞMB ð38Þ tities, i.e. the internal energy, the volume and the number of moles
376 A. Favache, D. Dochain / Journal of Process Control 19 (2009) 371–379

of each species [5]. Based on this, Glansdorff and Prigogine have 70

established a general evolution criterion for all macroscopic sys-
Tinit = 290 K
tem submitted to time independent boundary conditions [12]. This 60 Tinit = 310K

Entropy production [J/(K s)]

criterion states that there exist some flows of extensive quantities Tinit = 330 K
bJ i and some associated thermodynamic forces Y b i such that steadystate
50
X dY bi
W¼ bJ i 60 ð45Þ
i
dt 40

Note that the total entropy production rS can always be written as a

sum of some fluxes Ji and thermodynamic forces Yi : 30
X
rS ¼ Ji Y i ð46Þ
20
i

Yet the choice of the expression of these fluxes and thermodynamic

10
forces is not unique. It is also worth noting that in the most general 0 50 100 150 200 250
case, the fluxes bJ i and the thermodynamic forces Y b i defined by Glan-
Time [s]
sdorff and Prigogine are not fulfilling (46), i.e.:
X Fig. 3. Evolution of the entropy production for different initial temperature values.
bJ i Y
b i –rS ð47Þ
i

The equality is fulfilled only in special cases, e.g. for systems for A first possible alternative is to consider the quantity W defined
which one particular choice of Ji and Yi is such that they fulfill the by Glansdorff and Prigogine and stated in (45). If W was the varia-
Onsager reciprocal relations. For those systems, the thermodynamic tion of some function of the state along the trajectories, then this
fluxes appearing in the expression of the entropy production can be function would be an appropriate Lyapunov candidate. It can be
written as a linear combination of the thermodynamic forces: shown that for the CSTR the fluxes bJ i and the corresponding forces
X b i are given by following expressions:
Y
Ji ¼ Lij Y j ð48Þ
j
Flux bJ j bj
Force Y
where Lij is a symmetric matrix. The time variation of the entropy qC in in q 1
i ui  V U þ hðT w  TÞ T
production is then given by following expression: li
ðqC in q
i  V ni Þ T
drS X dJ i X dY i X dY i X dY i
rV
Aff
¼ Yi þ Ji ¼2 Lij Yj ¼ 2 Ji ð49Þ T
dt i
dt i
dt i;j
dt i
dt

It can be shown that for this special case, the entropy production For the CSTR the function W is unfortunately not equal to the
variation is negative, i.e.: variation of some function of the state along the trajectories. This
X dY i 1 drS has motivated the work of Tarbell [17] who looked for a positive
Ji ¼ 60 ð50Þ constant  such that W is the derivative of a function U(X) along
dt 2 dt
i the trajectories. But an appropriate factor  is found only when the
Comparing the above relation (50) to (45) shows that in this special equilibrium point is close to the chemical equilibrium, i.e. for the
b i ¼ Y i and W ¼ 1 drS . Therefore systems fulfilling the
case bJ i ¼ J i and Y 2 dt
highest conversion equilibrium point.
abovementioned phenomenological assumptions, the entropy pro- Let us now consider the work of Rouchon and Creff [15] on the
duction rS reaches a minimum at steady state and can be consid- stability analysis of the flash dynamics, for which the entropy pro-
ered as a Lyapunov function candidate. duction is a Lyapunov function although the flash dynamics do not
However a limited number of systems fulfill these constraints fulfill the Onsager relations. Unlike the CSTR model, the flash
and the CSTR unfortunately does not belong to this class of sys- dynamics has no multiple steady states. Furthermore numerical
tems. Indeed, from (17) and (46), the following fluxes and efforts simulations show that the stable equilibria of the CSTR are not lo-
can be deduced: cal minima of the entropy production (see Fig. 3). Let us therefore
investigate the differences between the flash and the CSTR from
 fluxes: J ¼ ½Fin _ the viewpoint of the entropy production variation.
X ; rV; Q
h T  T  iT
T
 driving forces: Y ¼ : ooXS jX  : ooXS jX in ; : oonS jX C; : oU
oS 1
jX  T ext The time derivative of rS can then be computed from (17) as
follows:
with C the matrix of the stoichiometric coefficients. If the
drS oT rS dX
Onsager relations were fulfilled, this would mean that the fluxes ¼
were linear functions of the driving forces, i.e. the molar inflow
dt oX dt
T     
of A would be a proportional to the difference of chemical potential d X o2 S dX out oS  1 oT Q_ dX oT S orn dX
¼ þ F þ  þ
of A between the inlet and the inside of the reactor, of the affinity dt oX 2 dt X
oU X T ext oX dt oX oX dt
of the reaction and of the temperature difference between the ð51Þ
reactor and the environment. This would be verified if the the inlet
flowrates were due to diffusion instead of convection. Therefore The Gibbs-Duhem equation that links the degree one homogeneity
the entropy production would not be a priori a good Lyapunov can- of the thermodynamical potentials and the Gibbs’ relation is written
didate, yet the above arguments might be a good source of infor- as follows :
mation for deriving an appropriate alternative Lyapunov function  
d oS o2 S dX
candidate. On Fig. 3, it can be seen that indeed the entropy produc- 0 ¼ XT ¼ XT 2 ð52Þ
tion does not decrease along the trajectories. dt oX oX dt
 A. Favache, D. Dochain / Journal of Process Control 19 (2009) 371–379 377

From (13), (52), the time derivative of rS becomes: Table 1 2

dD
 Sign of dteq
T  
drS d X o S dX oS  2
1 oT Q_ dX oT S orn dX dDeq dT dðD2eq Þ
¼ 2
þ   þ ð53Þ Deq(T) dT dt dt
dt dt oX dt oU X T ext oX dt oX oX dt
+ + + +
In the case of the flash, only the first term exists since there is no +  + 
+ +  
reaction and no heat exchange. The stability of the equilibrium de-
o2 S +   +
rives from the concavity of the entropy function (oX 2 < 0). For the  + + 
CSTR, if this result obviously still holds, we also have to account   + +
for the other two terms of the right hand side of (53). Unfortunately  +  +
it is not possible to deduce the sign of these two terms. In summary,    

we can say that the existence of the equilibrium point multiplicity

and of the instability of one of them is related on one hand on the
heat exchange and on the other hand on the chemical reaction. Table 1 gives its sign with respect to the sign of the other terms.
dD
While the sign of Deq(T) and dTeq are known since they only depend
4.3. Stability analysis based on the energy on T, the sign of dT
dt
depends both on T and nA. Simulation runs con-
firm that depending on the initial value of nA and nB, different
Since the entropy does not provide substantial arguments at behaviours of D2eq ðTÞ are observed, sometimes strictly decreasing,
this point for a physically based Lyapunov stability analysis, let sometimes increasing before reaching a maximum and then
us therefore concentrate on the internal energy and more precisely decreasing (i.e. the temperature is moving away from the equilib-
on the thermal energy, in line with the classical intuitive process rium value before going back to it). The function D2eq ðTÞ cannot
engineering arguments on the stability/instability of the three therefore be considered as a global Lyapunov function but it might
equilibrium points of the CSTR model. worth looking to what happens if the initial conditions of the triple
For this purpose, let us define the function Deq(T) from the right (nA,nB,T) are restricted to a limited set of the positive orthant. Let us
hand side of the dynamical equation of the temperature when first look to the subsets in which D2eq ðTÞ is decreasing and then ana-
nA = neq eq
A ðTÞ and nB = nB ðTÞ: lyze if these are invariant sets or if they contain such invariant sets
(with the objective to possibly apply the LaSalle theorem [13]).
Deq ðTÞ ¼ ½a  cT þ rðT; neq eq
A ðTÞ; nB ðTÞÞVðDr HÞðTÞ ð54Þ
First note that the reaction rðT; neq eq
A ðTÞ; nB ðTÞÞ is an increasing
with function of nA and a decreasing function of nB. This implies that
 
a ¼ hT w þ q C inA C vA þ C I C vI T in ; c ¼ h þ qðC inA C vA þ C I C vI Þ dT
Cv < Deq ðT Þ; if nA < neq
A ðTÞ and nB P neq
B ðTÞ ð57Þ
dt
and neq and
A ðTÞ neq
being the following function defined from the
B ðTÞ
dT
Cv > Deq ðTÞ; if nA > neq
A ðTÞ and nB 6 neq
B ðTÞ ð58Þ
equilibrium value of nA and nB (25): dt
8 eq in eq It is rather straightforward to see that D2eq ðTÞ is decreasing along the
< nB ðTÞ ¼ bðC A V  nA ðTÞÞ
>
state trajectories in the following domains:
q eq
ðkf ðTÞ þ V ÞnA ðTÞ ð55Þ
>
:
kb ðTÞ½bðC in eq b in X 1 ¼ ðT; nA ; nB ÞjT < T 1 ; nA > neq eq
A ðTÞ; nB 6 nB ðTÞ
A V  nA ðTÞ ¼ qC A
[ fðT; nA ; nB ÞjT max
An important feature of the function Deq(T) is that it is equal to zero
for the three equilibrium values of T. Therefore Deq(T)2, which is < T < T 3 ; nA > neq eq
A ðTÞ; nB 6 nB ðTÞg ð59Þ
represented on Fig. 4, has three local minima in T 1 , T 2 and T 3 .
The time derivative of Deq(T)2 is equal to: X 2 ¼ fðT; nA ; nB ÞjT 1 < T < T min ; nA < neq
A ðTÞ; nB

dðD2eq Þ dDeq dT P neq eq eq

B ðTÞg [ fðT; nA ; nB ÞjT > T 3 ; nA < nA ðTÞ; nB P nB ðTÞg ð60Þ
¼ 2Deq ðTÞ ð56Þ
dt dT dt
The union of X1 and X2 are indeed two disconnected sets, one
around the equilibrium ðn  B1 ; T 1 Þ, and the other around the equi-
 A1 ; n
librium ðn  B3 ; T 3 Þ, i.e. around the two stable equilibria.
 A3 ; n
x 1012 Let us denote by D1 and D3 each of these subsets (see Fig. 5). It is
worth noting that none of D1 and D3 are invariant sets. Indeed let
us consider the following boundaries:

oD1 ¼ fðT; nA ; nB ÞjT ¼ T 1 ; nA < neq eq

A ðTÞ; nB P nB ðTÞg ð61Þ
oD3 ¼ fðT; nA ; nB ÞjT ¼ T 3 ; nA < neq
A ðTÞ; nB P neq
B ðTÞg ð62Þ
Δ eq (T)

From (57), we know that on these boundaries:

2

dT
Cv < Deq ðTÞ ¼ 0
dt
and the time derivative points towards the exterior of the domain.
In order to find out invariants sets of the system or attraction
domains for the equilibria, one could draw the phase plane of the
system. More precisely, it is possible to split the space into sub-
0
T1 Tmin T2 Tmax T3 spaces delimited by the three surfaces given by
  
Temperature dxi
Si ¼ ðT; nA ; nB Þ ¼0 ; with i ¼ 1; 2; 3 and x ¼ ðT; nA ; nB Þ
dt
Fig. 4. D2eq ðTÞ.
378 A. Favache, D. Dochain / Journal of Process Control 19 (2009) 371–379

1800 0
eq eq n^eqA(T)
nA (T), n B (T)
D1 neqA (T)
Quantity of A [mol]

Quantity of B [mol]
1200 1200

nA [mol]
600 2400

D3

0 3600
T1 TminT2 Tmax T3
Temperature
0
Fig. 5. Shaded domains on the left (X1) and right (X2) of both equilibrium points T1 T_1 T_a T_2 T_b T_3
and T3 . température

Fig. 6. Trajectory directions in the CSTR.

In each of the delimited subspaces, the general orientation of the
tangent vector can be found since the surface Si divides R3 into
8 9
two parts: > T 6 T2 >
< =
      X 3 ¼ nA 6 maxðn ^ eq ðTÞÞ for 0 < T 6 T ð65Þ
dxi dxi > A max
>
ðT; nA ; nB Þ >0 and ðT; nA ; nB Þ <0 :
^ eq
;
dt dt nA 6 n A ðTÞ for T max < T 6 T 2
8 9
>
< T P T2 >
=
This allows to find invariant sets delimited by parts of
X4 ¼ nA P n^ eq ðTÞ for T 2 6 T 6 T min ð66Þ
Si ði ¼ 1; 2; 3Þ. >
:
A >
;
The exercise can become rapidly tedious for systems of dimen- nA P minðn ^ eq
A ðTÞÞ for T min < T
sion larger than two. As a matter of illustration and in order to keep
The above analysis has led us to identify invariant sets as illustrated
the approach simple enough, let us concentrate on the irreversible in Fig. 6 for the two dimensional system case. It is obvious that such
reaction case, which can be considered as a particular case of the
an approach is quite appealing and, with that respect, would de-
reversible reaction as mentioned at the end of Section 2. In this in-
serve to be possibly further explored. However, although this anal-
stance, the system dimension is equal to 2. In order to find out
ysis could lead to invariants sets for the stable equilibrium points of
invariants sets of the system or attraction domains for the equilib-
the CSTR, the link with the physics is unfortunately not obvious.
ria, let us draw the phase plane of the system. More precisely let us
split the plane (nA,T) into different zones in which the time varia-
^ eq 5. Conclusion
tion of nA and T are analyzed. Let us define n A ðTÞ as the value of
nA for which the time derivative of T is equal to zero, i.e.:
This paper was dedicated to the stability analysis of a typical
cT ¼ rðT; n^eq process example, the CSTR, by considering thermodynamics and
A ÞVðDr HÞ þ a ð63Þ
energy considerations. Unlike for electrical and mechanical sys-
The curve representing n ^ eq
A ðTÞ obviously crosses the curve repre-
tems, the link between Lyapunov stability theory and energy of
eq
^ eq
senting nA ðTÞ. The time derivative of n A ðTÞ is given by the follow-
the system is not straightforward. Different Lyapunov function
2
ing expression : candidates related to the thermodynamics, more specifically based
on the entropy, the entropy production and the internal energy, are
^eq
dnA
c  VðDr HÞ oT
or
considered. These provide new insight and physical interpretation
¼ or
ð64Þ
dT V onA ðDr HÞ in the stability/instability of the equilibrium points of the CSTR. In
particular, the local stability analysis based on the first Lyapunov
For low and high temperatures, oT or ^ eq
is small so that n A ðTÞ is increas- method has resulted in an extension of the results of [11] for less
ing for low and high temperatures, and decreasing for intermediate restrictive conditions on the system dynamics and thermodynam-
temperatures. Similarly to the fact that the curve neq A ðTÞ is delimit- ics. Besides invariants sets for the stable equilibrium points of the
ing in the plane (nA,T) the zones where the quantity of A is increas- CSTR can be identified by considering an internal energy based
ing or decreasing, the curve n ^ eq
A ðTÞ is delimiting the zones of Lyapunov function. Yet thermodynamic functions like the entropy
increasing or decreasing temperature. The decreasing temperature and the entropy production as candidate Lyapunov functions exhi-
zones are under the curve n ^ eq
A ðTÞ while the decreasing nA zones bit their limitation in the case of open reaction systems. It has to be
eq
are above nA ðTÞ. This is schematically represented in Fig. 6. pointed out that the link between Lyapunov stability theory and
From Fig. 6, we can identify two invariant sets around each sta- the energy of reaction systems is not fully solved nor undestood,
ble equilibrium point. If we note by Tmax (resp., Tmin) the tempera- and that there remain open questions that need to be addressed
ture corresponding to a local maximum (resp., minimum) of n ^ eq
A ðTÞ, before obtaining a fully comprehensive thermodynamics based
both invariants are delimited by the following curves: Lyapunov stability theory for reaction systems.

Acknowledgements
2
An explicit expression can be given for mass action law kinetics: This paper presents research results of the Belgian Network
^ eq
cT  a DYSCO (Dynamical Systems, Control, and Optimization), funded
n A ðTÞ ¼
kðTÞVðDr HÞ by the Interuniversity Attraction Poles Programme, initiated by
 A. Favache, D. Dochain / Journal of Process Control 19 (2009) 371–379 379

the Belgian State, Science Policy Office. The scientific responsibility controllability of the continuous stirred tank reactors, Chem. Engrg. Sci. 29
(1974) 1917–1926.
rests with its authors. Audrey Favache is a fellow student of the
[10] G.R. Gavalas, Nonlinear Differential Equations of Chemically Reacting Systems,
Belgian Fonds National de la Recherche Scientifique (FNRS). Springer Verlag, Berlin, 1968.
[11] C. Georgakis, On the use of extensive variables in process dynamics and
References control, Chem. Engrg. Sci. 41 (6) (1986) 1471–1484.
[12] P. Glansdorff, I. Prigogine, On a general evolution criterion in macroscopic
physics, Physica 30 (1964) 351–374.
[1] A.A. Alonso, B.E. Ydstie, Stabilization of distributed systems using irreversible [13] H.K. Khalil, Nonlinear Systems, MacMillan, New York, 1992.
thermodynamics, Automatica 37 (2001) 1739–1755. [14] W.H. Ray, New approaches to the dynamics of nonlinear system with
[2] A.A. Alonso, B.E. Ydstie, J.R. Banga, From irreversible thermodynamics to a implications for process and control system design, in: T.F. Edgar, D.E.
robust control theory for distributed process systems, J. Process Control 12 Seborg (Eds.), Proceedings of CPC II, 1981.
(2002) 507–517. [15] P. Rouchon, Y. Creff, Geometry of the flash dynamics, Chem. Engrg. Sci. 48 (18)
[3] L.T. Antelo, I. Otero-Muras, J.R. Banga, A.A. Alonso, A systematic approach to (1993) 3141–3147.
plant-wide control based on thermodynamics, Comput. Chem. Engrg. 31 [16] S.I. Sandler, Chemical and Engineering Thermodynamics, John Wiley, New
(2007) 677–691. York, 1999.
[4] R. Aris, Elementary Chemical Reactor Analysis, Dover Publications, New York, [17] J.M. Tarbell, A thermodynamic Lyapunov function for the near equilibrium
2000. CSTR, Chem. Engrg. Sci. 32 (1977) 1471–1476.
[5] H.B. Callen, Thermodynamics and An Introduction to Thermostatics, John [18] R.B. Warden, R. Aris, N.R. Amundson, An analysis of chemical reactor stability
Wiley, New York, 1985. and control – VIII. The direct method of Lyapunov. Introduction and
[6] W.R. Dammers, M. Tels, Thermodynamic stability and entropy production in applications to simple reactions in stirred vessels, Chem. Engrg. Sci. 19 (3)
adabiatic stirred flow reactors, Chem. Engrg. Sci. 29 (1974) 83–90. (1964) 149–172.
[7] S.R. de Groot, P. Mazur, Non-Equilibrium Thermodynamics, Dover Publications, [19] J. Willems, Stability Theory of Dynamical Systems, Nelson, New York, 1970.
New York, 1984. [20] B.E. Ydstie, A.A. Alonso, Process systems and passivity via the Clausius-Planck
[8] M. Feinberg, Lectures on Chemical Reaction Networks, University of inequality, Syst. Control Lett. 30 (1997) 253–264.
Wisconsin-Madison, Mathematics Research Center, 1979. [21] B.E. Ydstie, Passivity based control via the second law, Comput. Chem. Engrg.
[9] M. Fjeld, O.A. Asbjornsen, K.J. Astrom, Reaction invariants and their 26 (2002) 1037–1048.
importance in the analysis of eigenvectors, state observability and