Mass Balan

e Modelling of Biopro esses

Olivier Bernard

COMORE, INRIA, Fran e

Le tures given at the

Summer S hool on Mathemati al Control Theory
Trieste, 3-28 September 2001

LNS0280012

 obernardsophia.inria.fr
 Abstra t
These le ture notes detail how to design a model for a biologi al
pro ess. The diÆ ulty is due to the fa t that, on the ontrary to other
elds (me hani s, ele troni s, et .) there does not exist any validated
law to des ribe the behaviour of biologi al systems. Nevertheless, these
systems satisfy the mass onservation prin iple. On this basis, the
le ture explains how to derive a model whi h will represent the main
mass transfer within the system. The method onsists of three steps.
First determine the rea tion s heme and dene a model in whi h the
mi robial kineti s are not spe ied. Then nd an analyti al expression
for the biologi al kineti s. Finally validate the model trying to test
separately the dierent hypotheses assumed during model design.
Contents

1 Introdu tion 773

2 Prin iple of a biorea tor 774
2.1 The use of mi roorganisms . . . . . . . . . . . . . . . . . . . . 774
2.2 The main types of biorea tors . . . . . . . . . . . . . . . . . . 775
2.3 Working of a biorea tor . . . . . . . . . . . . . . . . . . . . . 776
2.3.1 Presentation . . . . . . . . . . . . . . . . . . . . . . . 776
2.3.2 Bat h mode . . . . . . . . . . . . . . . . . . . . . . . 776
2.3.3 Fedbat h mode . . . . . . . . . . . . . . . . . . . . . . 776
2.3.4 The ontinuous mode ( hemostat) . . . . . . . . . . . 778
2.3.5 The Sequen ing Bat h Rea tors (SBR) . . . . . . . . . 778
3 The mass balan e modelling 778
3.1 Introdu tion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 778
3.2 Rea tion s heme . . . . . . . . . . . . . . . . . . . . . . . . . 780
3.3 Choi e of the rea tions and of the variables . . . . . . . . . . 781
3.4 Example 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 781
4 The mass balan e models 782
4.1 Introdu tion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 782
4.2 Example 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 782
4.3 Matrix representation . . . . . . . . . . . . . . . . . . . . . . 784
4.3.1 Example 2 ( ontinued) . . . . . . . . . . . . . . . . . . 784
4.3.2 Example 1 ( ontinued) . . . . . . . . . . . . . . . . . . 785
4.4 The gaseous ows . . . . . . . . . . . . . . . . . . . . . . . . . 785
4.5 Ele tro neutrality and aÆnity onstants . . . . . . . . . . . . 786
4.6 Example 1 ( ontinued) . . . . . . . . . . . . . . . . . . . . . . 786
4.6.1 Gaseous ows . . . . . . . . . . . . . . . . . . . . . . . 786
4.6.2 AÆnity onstants . . . . . . . . . . . . . . . . . . . . . 787
4.6.3 Ele tro neutrality of the solution . . . . . . . . . . . . 787
4.6.4 Con lusion . . . . . . . . . . . . . . . . . . . . . . . . 787
4.7 Con lusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 788
5 Modelling of the kineti s 789
5.1 Introdu tion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 789
5.2 The mathemati al onstraints . . . . . . . . . . . . . . . . . . 789
5.2.1 Positivity of the variables . . . . . . . . . . . . . . . . 789
 5.2.2 Variables that are ne essary for the rea tion . . . . . . 790
5.2.3 Example 1 ( ontinued) . . . . . . . . . . . . . . . . . . 790
5.2.4 Phenomenologi al knowledge . . . . . . . . . . . . . . 791
5.3 The growth rate . . . . . . . . . . . . . . . . . . . . . . . . . 791
5.3.1 The Monod model . . . . . . . . . . . . . . . . . . . . 791
5.3.2 Haldane model . . . . . . . . . . . . . . . . . . . . . . 792
5.3.3 Multiple limitations . . . . . . . . . . . . . . . . . . . 792
5.4 Kineti s representation using neural networks . . . . . . . . 792
6 Model validation 794
6.1 Introdu tion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 794
6.2 Validation of the rea tion s heme . . . . . . . . . . . . . . . . 794
6.2.1 Mathemati al prin iple . . . . . . . . . . . . . . . . . 794
6.2.2 Example 4 . . . . . . . . . . . . . . . . . . . . . . . . . 795
6.3 Qualitative model validation . . . . . . . . . . . . . . . . . . 797
6.3.1 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 798
6.4 Global model validation . . . . . . . . . . . . . . . . . . . . . 798
6.4.1 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 800
7 Mass balan e models properties 801
7.1 Boundness and positivity of the variables . . . . . . . . . . . 801
7.2 Equilibrium point and lo al behaviour . . . . . . . . . . . . . 804
7.2.1 Introdu tion . . . . . . . . . . . . . . . . . . . . . . . 804
7.2.2 Equilibrium points and lo al stability . . . . . . . . . 804
7.2.3 Global behaviour . . . . . . . . . . . . . . . . . . . . . 805
7.2.4 Asymptoti behaviour . . . . . . . . . . . . . . . . . . 805
7.2.5 Example 4 ( ontinued) . . . . . . . . . . . . . . . . . . 805
8 Con lusion 806
Referen es 810
 Mass Balan e Modelling of Biopro esses 773

1 Introdu tion

System modelling in general is diÆ ult and requires time to properly under-
stand the system and identify a model. This exer ise is ompli ated when
the system integrates living organisms. On the ontrary to domains like
physi s where laws that are known sin e enturies (Ohm law, ideal gas re-
lationship, fundamental prin iple in me hani s, thermodynami prin iple,
...) an apply, most of the biologi al models rely on empiri al mathemati al
expressions. These laws result from a priori ideas on the working of the
system (metabolism, trophi relationships, et .) or, in some rare ases, have
been estimated from some experiments. Sin e it is not possible to use laws
that are admitted by everybody and that have been extensively validated
and used, it is primordial to hara terise the reliability of the mathemati-
al expressions used during the model development. This implies that the
reliability of the used relationships must be sorted hierar hi ally during the
model development. In this hapter, we will see how to organise the knowl-
edge in the model in order to distinguish a reliable part issued from the
mass balan e and a more spe ulative part whi h will represent the ba terial
kineti s.
The model quality and the model stru ture must above all be determined
with respe t to the model obje tives. Indeed, a model an be developed for
very dierent purposes that must be learly identied. Will the model be
used in order to:
 Reprodu e an observed behaviour
 Explain an observed behaviour
 Predi t the system evolution
 Understand some of the system me hanisms
 Estimate non measured variables
 Estimate pro ess parameters
 A t on a system to regulate and impose the values for its variables
 Dete t anomalies in the pro ess working
 ...
774 O. Bernard

Depending on the modelling obje tives and resour es, a formalism must
be hosen. If the spatial heterogeneity is important and must be taken into
a ount in the model, a parameter distributed model must be written (using
e.g. partial dierential equations). If the modelling aims at the improvement
of a metabolite produ tion during transient phases, the system dynami s
must be represented in the model.
Moreover, besides its obje tives, the model must also be in adequation
with the available data. Indeed a omplex model involving a large number of
parameters will also require a large amount of data to identify its parameters
and to validate the model.
Finally, if we remember that most of the laws used in biology are spe u-
lative, the key step in the modelling of biopro esses is the model validation.
This step is often negle ted, despite its determinant role to guaranty the
model quality. In parti ular it is ru ial to demonstrate that the model
rea hes properly the goals for whi h it was developed.

2 Prin iple of a biorea tor

2.1 The use of mi roorganisms

The fermentation prin iple onsists in exploiting metaboli rea tions that
take pla e in the ell of a mi ro-organism (ba teria, yeast, phytoplankton,
et .). In order to a tivate the mi ro-organisms interesting metaboli path-
ways, some spe i environmental onditions must be applied (temperature,
pH, nutrient on entration). The mi roorganisms generally need nutrients
to growth and pre ursors or a tivators in order to produ e spe i mole ules.
The simplest required rea tion is the growth pro ess itself in order to re over
the biomass of mi roorganisms.
In these metaboli rea tions, we an distinguish the following bio hemi al
omponents:
 the substrates Si, whi h are ne essary for the goal of the fermentation
(growth of the mi roorganisms and/or pre ursor for the metabolite
to be produ ed). The substrate asso iated with growth must ontain
all the elements ne essary to sustain growth (i.e. N, C, K, P, Fe,
...). In general, these elements are added in ex ess so that they are
never limiting during the ultivation. Only the main nutrients ( arbon,
nitrogen or phosphorus sour e) are monitored along the ultivation.
 Mass Balan e Modelling of Biopro esses 775

 mi robial biomasses (denoted Xi ). The mi roorganisms an be of var-

ious type and spe ies (ba teria, phytoplankton, fungi, yeast, et .);
 the produ ts of the bio hemi al rea tions,(denoted Pi ). These prod-
u ts an be in the agro-industrial eld ( heese, beer, wine, ... ),
hemistry (enzymes, olourings...), pharma euti al industry (antibi-
oti s, hormones, vitamins...) or for energy produ tion (ethanol, bio-
gas...)...
 atalysts: they an neither be produ ed nor onsumed during the re-
a tion, but they are ne essary.
Depending on the obje tives of the fermentation, spe i mi roorganisms
will be grown in order to enhan e:
 produ tion of biomass itself. It is for example the ase for the produ -
tion of ba ker yeast.
 produ tion of a metabolite. The goal is to enhan e the ellular syn-
thesis of a parti ular ompounds (ethanol, peni illin, ...).
 substrate uptake. In this ase, the substrate degradation itself is the
obje tive. This is more spe ially used to remove pollutants from a
liquid medium. Most of the biologi al depollution pro esses are among
this ategory.
 phenomenologi al studies. In this parti ular ase the fermentation
aims a better knowledge of the mi roorganism. The appli ation an
be to better understand how the mi roorganisms grow in the natural
eld.

2.2 The main types of biorea tors

There are a great deal of dierent biorea tors. Depending on the type of
mi roorganisms that are grown, they will need a support to settle or an be
free in the liquid. They an resist to more or less intense shearing onstraints
whi h will impli ate a spe i steering system. These two main requirements
will determine the type of biorea tor. Two lasses an be identied [1℄:
 stirred tank rea tors (CSTR) in whi h the medium is homogeneous
and ea h element of volume will represent the on entrations in the
whole fermenter
776 O. Bernard

 the biorea tors with non homogeneous on entration along spa e. In

parti ular the biorea tor for mi roorganisms using a support to growth
( alled a \bed") are in this ategory.
When the medium is homogeneous it an be des ribed by ordinary dif-
ferential equations. When a strong spatial distribution must be taken into
a ount a model based on partial dierential equations are more appropriate.
In this le ture we will present only the CSTR modelled with ODE.

2.3 Working of a biorea tor

2.3.1 Presentation
Figure 1 presents a simplied on eptual s heme explaining the prin iple of
a biorea tor. It is mainly a ulture vessel of volume V where the mi roor-
ganisms grow. A pipe feeds the vessel with an in uent medium (with ow
rate Qin ) and another one withdraws the ulture medium with a ow rate
Qout .
Depending on the way the fermenter is fed and withdrawn, 3 basi work-
ing modes an be identied (gure 2).

2.3.2 Bat h mode

The system is in bat h during the fermentation, and has a onstant volume,
sin e no feeding or withdrawal are performed during the fermentation. An
ino ulum of mi ro-organisms is introdu ed at the initial time with all the
nutrients and substrates. The biomass or the nal produ t are re overed at
the end of the fermentation. The advantage of this approa h is that it avoids
the ontaminations with other ba teria that an ome in an open system.
The drawba k is the limited means of a tion to a t on the fermentation
(pH, temperature, aeration...). Therefore the bat h mode is often the less
optimal from the automati ontrol point of view to optimise a ost riterion.
Nevertheless, this is the most used mode in the industry.

2.3.3 Fedbat h mode

As for the bat h mode the duration of a fedbat h is nite. But here the
fermenter is fed and starts from a volume V0 to rea h a volume Vf at the
end of the fermentation. This mode allows a better ontrol of the growth and
 Mass Balan e Modelling of Biopro esses 777

Qin Qout

Sin
111
000
000
111
000
111

Substrate

Biomass

Product

11111
00000
00000
11111
00000
11111
00000
11111
0
1
00000
11111
00000
11111
0
1
00000
11111
00000
11111
0
1
0
1
0
1
0
1
0
1

Figure 1: Prin iple of a biorea tor

1111111
0000000
1111111
0000000 0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111

111111111
000000000
000000000
111111111 111111111
000000000
000000000
111111111
000000000
111111111 000000000
111111111
000000000
111111111
000000000
111111111 000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111 000000000
111111111
000000000
111111111
000000000
111111111 000000000
111111111
000000000
111111111 000000000
111111111
000000000
111111111
000000000
111111111 000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111 000000000
111111111
000000000
111111111
000000000
111111111 000000000
111111111
000000000
111111111
000000000
111111111 000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111 000000000
111111111
000000000
111111111
000000000
111111111 111111111
000000000 000000000
111111111
000000000
111111111
000000000
111111111 000000000
111111111 000000000
111111111
000000000
111111111
000000000
111111111 000000000
111111111
000000000
111111111 000000000
111111111
00000
11111
000000000
111111111 000000000
111111111 000000000
111111111
00000
11111
000000000
111111111 000000000
111111111 000000000
111111111
00000
11111
000000000
111111111 000000000
111111111
00000
11111
00000
11111
00000
11111

Batch Fedbatch Chemostat

Figure 2: The various working modes of the biorea tors

778 O. Bernard

biotransformation pro ess along the fermentation. The fedbat h pro esses
are often in losed loop. This operating mode is parti ularly used when the
produ t to be re overed ne essitates to empty the biorea tor like e.g. for
intra ellular omponents.

2.3.4 The ontinuous mode ( hemostat)

This is the most popular working mode in the eld of wastewater treat-
ment. The volume of the biorea tor is onstant sin e the in uent ow rate
is equal to the euent ow rate. This mode provides the ri hest dynami s,
and therefore presents the more latitude to optimise the pro ess. It is also
often used in laboratories to study the physiology of a mi roorganism. The
advantage is also that it allows important produ tions in small size rea tors.

2.3.5 The Sequen ing Bat h Rea tors (SBR)

It is a ombination of the previous working mode. The idea is to re over the
biomass before emptying the biorea tor. For this, the agitation is stopped
to let the biomass settle. The dierent steps used for wastewater treatment
are presented on Figure 3.
In the same way, the SFBR (sequen ing fedbat h rea tor) is a SBR with
a stage of lling that follows a fedbat h mode.

3 The mass balan e modelling

3.1 Introdu tion

The modelling of biologi al systems is deli ate be ause it is not based on val-
idated laws, like in other elds (me hani s, ele troni s, et .). The evolution
of mi roorganisms is very omplex and does not follow any lear law. Never-
theless, this system has to respe t some rules, like all the physi al systems.
For example, the mass onservation, the ele tro neutrality of the solutions,
et . We will see in this se tion how to take these aspe ts into a ount in
the model design. As a result, this mass balan e approa h will guaranty a
ertain robustness in the model.
 Mass Balan e Modelling of Biopro esses 779

111111
000000
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
0000000000
1111111111
1111111111
0000000000
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
0000000000
1111111111
1111111111
0000000000 0000000000
1111111111
0000000000
1111111111 0000000000
1111111111 1111111111
0000000000
0000000000
1111111111
0000000000
1111111111 0000000000
1111111111 0000000000
1111111111
0000000000
1111111111 0000000000
1111111111
0000000000
1111111111 0000000000
1111111111
0000000000
1111111111

1−Filling 2− Reaction 3−Settling

111111
000000
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111

1111111111
0000000000
0000000000
1111111111 0000000000
1111111111
0000000000
1111111111
1111111111
0000000000
0000000000
1111111111 111111111
000000000 0000000000
1111111111
000
111 000000000
111111111 0000000000
1111111111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
4−Sludge removal
000
111 5−Emptying 6−Back to step 1 ...

Figure 3: SBR (sequen ing bat h rea tors): representation of the dierent steps
780 O. Bernard

3.2 Rea tion s heme

The rea tion s heme of a bio hemi al pro ess is a ma ros opi des ription of
the set of biologi al and hemi al rea tions whi h represents the main mass
transfer within the fermenter. A formalism lose to this used in hemistry is
adopted [2℄. A set of substrates Si are transformed into produ ts Pi following
3 possibilities:
 The rea tion is a pure hemi al rea tion, and no biomass is involved.
The rea tion is then a lassi al hemi al rea tion:
S1 + S2 + ::: + Sp ! P1 + ::: + Pq
 The rea tion is atalysed by a biomass X . The biomass a ts only as
a atalyser and the rea tion is not asso iated with the growth of the
mi roorganisms:
S1 + S2 + ::: + Sp X! P1 + ::: + Pq
 The rea tion is asso iated with growth of the mi roorganisms. There-
fore the biomass is also a produ t of the rea tion.
S1 + S2 + ::: + Sp X! P1 + ::: + Pq + X
The rea tion s heme is a on ise way to summarise at the ma ros opi
level a set of rea tions that are assumed to determine the pro ess dynami s.
The rea tion s heme is therefore based on the assumptions related to the
available phenomenologi al knowledge of the pro ess.
In general only the main omponents of a rea tion are represented. In-
deed, it would be very diÆ ult to present a real rea tion for the growth of a
mi ro-organism sin e a great deal of omponents are ne essary (Fe, Pb, F,
...).
In the sequel, we will detail the rea tion s heme by adding the yield
oeÆ ients asso iated with the onsumption (ki ) or the produ tion (ki0 ) of
ea h oeÆ ient. Moreover, we will also indi ate the rate of the rea tion ':
k1 S1 + k2 S2 + ::: + kp Sp '! k10 P1 + ::: + kq0 Pq + X
The onsumption rate of Si is thus ki ', the produ tion rate Pi is thus ki0 '.
By onvention ' orresponds to the produ tion rate of the biomass.
In the sequel we will assume that the rea tion s heme is omposed of
a set of k biologi al or hemi al rea tions. We will onsidered n variables
( hemi al on entrations, biomass,...).
 Mass Balan e Modelling of Biopro esses 781

3.3 Choi e of the rea tions and of the variables

The hoi e of the number of rea tions to be taken into a ount and the
hoi e of the state variables is apital for the modelling purpose. It will be
guided by the available knowledge on the rea tion s heme on the basis of
the available data set. Often the omplexity of the model is too high with
respe t to the amount of data that are available to test and validate the
model. It must be hosen with parsimony, keeping in mind the obje tives of
the model.
The hoi e of the rea tions and of the variables will mainly determine
the model stru ture, it must be onsidered with are. We will see in se tion
6 how to validate this rea tion s heme.
We brie y present in Appendix A a pro edure to determine the number
of rea tions that must be taken into a ount with respe t to the available
data.
In the sequel, we will assume that the rea tion s heme:
 represents the main mass and ow repartition between the set of rea -
tions that intervene in the pro ess,
 is a set of rea tions whose yield oeÆ ients are onstant.

3.4 Example 1
We will onsider here the example of anaerobi digestion. This pro ess is
used to remove a polluting substrate (S1 ) from wastewater thanks to anaer-
obi ba teria. In fa t, this is a very omplex pro ess whi h involves several
dierent ba terial populations [3℄. If the modelling obje tive is to ontrol
this intri ate e osystem in order to improve the pollution removal, then we
need a rather simple model. This is why, to limit the model omplexity, we
onsider only two main ba terial populations. We assume therefore that the
dynami s an be des ribed by two main steps:
 An a idogenesis step (with a rate r (:)) in whi h the substrate S is
1 1
degraded by a idogeni ba teria (X1 ) and is transformed into volatile
fatty a ids (VFA) (S2 ) and CO2 :

k1 S1 1! X1 + k2 S2 + k4 CO2
r (:)
(1)
782 O. Bernard

 A methanogenesis step (with a rate r (:)), where the volatile fatty a ids
2
are degraded into CH4 and CO2 by methanogeni ba teria (X2 ).
k3 S2 2! X2 ; + k5 CO2 + k6 CH4
r (:)
(2)
The onstants k1 ; k2 ; k4 , respe tively represent the stoi hiometri oef-
 ients asso iated with substrate S1 onsumption, produ tion of VFA and
CO2 during a idogenesis. k3 ; k5 and k6 respe tively represent stoi hiomet-
ri oeÆ ients asso iated with VFA onsumption and with CO2 and CH4
produ tion during methanogenesis.
It is worth noting that in some sense this rea tion s heme has no biolog-
i al reality sin e biomasses X1 and X2 represent a set of dierent spe ies. In
the same way for substrates S1 and S2 whi h gathers a set of heterogeneous
ompounds. A lot of models an be found in the literature for this pro ess
[4, 3, 5℄. Generally, the des ription of the pro esses within the biorea tor are
mu h more detailed [6, 7℄ but it leads to models diÆ ult to use for ontrol
purpose.

4 The mass balan e models

4.1 Introdu tion

We will onsider a ontinuously stirred tank rea tor that guarantees a per-
fe t mixing. We will see that independently of the working mode (bat h,
fedbat h, ontinuous), the dynami al behaviour of the biologi al or hem-
i al ompounds in the rea tors an be dire tly dedu ed from the rea tion
s heme.
We will show on a very simple example how the dynami al model an be
established.

4.2 Example 2
We will onsider here the very simple example of the growth of a mi ro-
organism X on a substrate S with rate r(:):
r(:)
kS !X
The yield oeÆ ient asso iated with substrate onsumption is denoted
k.
 Mass Balan e Modelling of Biopro esses 783

We assume that the in uent ow rate is Qin and that the euent ow
rate is Qout . We denote by x and s the total amount of biomass and substrate
in the volume V of the biorea tor.
Let us onsider the evolution of V (t), x(t) and s(t) between two very
lose time instants t and t + dt.
The evolution of the total liquid volume V is rather simple:
V (t + dt) = V (t) + Qin dt Qout dt
For the biomass, we have to take into a ount the new biomass produ ed
between t and t + dt. The produ tion term in the whole volume V is r(:)V dt,
and thus: x
x(t + dt) = x(t) + r(:)V dt Qout dt
V
Note that, in order to ompute the biomass lost in the euent (in the volume
Qout dt) we assume that the on entration in the small volume is the same as
in the whole biorea tor (i.e. Vx ). At this point the hypothesis of homogeneity
in the rea tor is ru ial.
In the same way, for the substrate, we must also onsider the quantity of
substrate (with on entration Sin ) arriving between the two time instant:
s
s(t + dt) = s(t) + QinSin kr(:)V dt Qout dt
V
For a very small dt, we an then derive the following equations:
8 dx x
>
>
> = r(:)V Qout (3)
>
> dt
< ds V
s
> dt
= kr ( : ) V + Qin S in Qout
V
(4)
>
>
>
>
: dV
dt
= Qin Qout (5)
Now, let us rewrite this model in term of on entration i.e. using the
variables X = Vx and S = Vs ). It is straightforward to see that we get the
following model:
dX = r (:) DX
dt
dS = kr (:) + D (S
dt in S) (6)
dV = Q
dt in Qout
where D = QVin orresponds to the dilution rate.
Model (6) simplies for the various working modes:
784 O. Bernard

 Bat h. In this ase we have Qin = Qout = 0. The volume is then

onstant.
 Fed bat h. Here Qout = 0; dVdt = Qin, V is in reasing.
 Continuous mode. The volume V is onstant sin e Qin = Qout.
For sake of simpli ity, in the sequel we will not des ribe the fed bat h ase
and we will on entrate on the bat h or ontinuous mode. This simplies
the equation sin e we do not need the equation whi h fore asts the volume
evolution.

4.3 Matrix representation

The rea tion s heme leads to the following mass balan e model whi h de-
s ribes equivalently the mass ows within the biorea tor [2℄:
_ = Kr(:) + D(in  ) Q( ) (7)
Where  is the state ve tor ontaining all the pro ess ompounds and biomasses,
in is the ve tor of the in uent on entrations, r(:) is a ve tor of rea tion
rates. The matrix K ontains the stoi hiometri oeÆ ients (yields). Q( ),
represents the gaseous terms of ex hange between the liquid and the gas
phase. The dilution rate, D, is the ratio between the in uent ow rate Qin
and the rea tor volume V .
Remark 1 In the ase of the fed bat h pro ess, the state ve tor must also
ontain the volume V of the rea tor. The last equation will des ribe the
volume evolution ( f. equation (5)).

4.3.1 Example 2 ( ontinued)

Let us onsider model (6) working in ontinuous mode (V is onstant, D =
Qin ). The model an be rewritten as follows:
V
0 1 0 1 00 1 0 11
X
 A= 1 A r(:) + D  0 A  X AA
S k Sin S
It orresponds exa tly to the general model, (7) with:
0 1 0 1 0 1 0 1
X
= A ; K=
1 A ; in =  0 A ; Q( ) =  0 A
S k Sin 0
 Mass Balan e Modelling of Biopro esses 785

4.3.2 Example 1 ( ontinued)

Now let us ome ba k to the anaerobi digestion example (see se tion 3.4).
We will assume that the methane solubility is very low and therefore that it
dire tly goes into the gas phase. The arbon dioxide is stored in the liquid
phase where he enters in the inorgani arbon ompartment (C ).
The mass balan e model is then the following:
dX1
dt
= r1 (:) DX1 (8)
dX2
dt
= r2 (:) DX2 (9)
dS1
dt
= D(S1in S1 ) k1 r1 (:) (10)
dS2
dt
= D(S2in S2 ) + k2 r1 (:) k3 r2 (:) (11)
dC
dt
= D(Cin C ) qC ( ) + k4 r1 (:) + k5 r2 (:) (12)
where S1in , S2in and Cin are respe tively the in uent on entrations of sub-
strate, VFA and dissolved inorgani arbon. The term qC ( ) represents the
inorgani arbon ow rate (of CO2 ) from the liquid phase to the gaseous
phase.

4.4 The gaseous ows

In order to derive the mass balan e, we must take into a ount the om-
pounds whi h have a gaseous phase. Indeed, the gaseous spe ies an es ape
the biorea tor after going from the liquid to the gaseous phase (they an
also enter into the biorea tor).
We use for this Henry's law whi h des ribes the molar ow rate of a
ompound C from its liquid phase to its gaseous phase:
q = KL a(C C ?) (13)
Remark 2 If q < 0, it means that the gaseous ow will take pla e from the
gaseous phase to the liquid phase.
The transfer oeÆ ient KL a (1/T) highly depends on the operating on-
ditions and espe ially from stirring, and the ex hange area between the liq-
uid and the gaseous phases (size of the bubbles)[8, 1℄. The modelling of this
parameter with respe t to the operating onditions an be very deli ate.
786 O. Bernard

The quantity C ? is the saturation on entration of dissolved C . This

quantity is related to the partial pressure of gaseous C (PC ) thanks to
Henry's onstant:
C ? = K H PC (14)
Henry's onstant an also vary with respe t to the ompounds in the ulture
medium or the temperature.
Moreover, when several gaseous spe ies are simultaneously in the gaseous
phase, they must follow the ideal gas law. This will give a relationship of
onstant ratio between molar ow rates and partial pressures. For m gaseous
spe ies C1 : : : Cm :
P 1 P 2 P
= = : : : = m (15)
q 1 q 2 q m

4.5 Ele tro neutrality and aÆnity onstants

The ele tro neutrality of the solutions is a se ond rule that the biologi al
systems must respe t: the anions on entrations weighted by the number
of ele tri al harges must equal the on entration of ations with the same
weighting.
The hemi al rea tions are often well known and an aÆnity onstant
is generally asso iated. This onstant is generally related to the protons
on entration H + , and therefore to pH.

4.6 Example 1 ( ontinued)

4.6.1 Gaseous ows
The methane ow rate is dire tly related to methanogenesis:
qM = k6 r2 (:) (16)
The gaseous CO2 ow rate follows Henry's law:
qC ( ) = KL a(CO2 KH PC ) (17)
where PC is the CO2 partial pressure.
 Mass Balan e Modelling of Biopro esses 787

4.6.2 AÆnity onstants

In the anaerobi digestion example, we will use the ele tro neutrality and
the hemi al aÆnity onstants:
In the usual operating range of pH for these pro esses (6  pH  8) we
assume that the VFA are under their ionised form. The dissolved CO2 is in
equilibrium with bi arbonate:
CO2 + H2 O $ HCO3 + H +
The aÆnity onstant of this rea tion is then
HCO3 H +
Kb = (18)
CO2
4.6.3 Ele tro neutrality of the solution
The ations (Z ), are mainly ions whi h are not ae ted by bio hemi al
rea tions (Na+ ,...). Therefore, their dynami s will simply follow, without
modi ation the ation on entration Zin in the in uent, so that:
dZ
dt
= D(Zin Z ) (19)
The anions are mainly represented by the VFA and the bi arbonate. Ele tro
neutrality ensures then that:
Z = S2 + HCO3 (20)

4.6.4 Con lusion

If we add equation (19), the model an nally be rewritten under the matrix
form (7), with :
2 3 2 3
66
X1
77 66
1 0 77
66 X2 77 2 3 66 0 1 77
6 Z 77 r (:) 6 0 0 77
 = 666 77 ; r(:) = 4 1 5 ; K = 666 77 (21)
66 S1 77 r2 (:) 66 k1 0 77
64 S2 75 64 k2 k3 75
C k4 k5
788 O. Bernard

2 3 2 3
66
0 7 66
0 77
66 0 77 66 0 77
6 77 6 77
Zin 0
in = 666 77 ;
77 Q = 666 77 ; (22)
66 S1in 66 0 77
64 77 64 75
S2in 5 0
Cin qC ( )
An elimination of variables HCO3 , CO2 , and PC using equations (17),(15)
(18) and (20), leads to the following expression for PC ( ) ( f [9℄):
p 2
 4KH PT (C + S2 Z )

PC ( ) = (23)
2KH
setting:  = C + S2 Z + KH PT + kk6a r2 (:), we nally get
L

qC ( ) = kL a(C + S2 Z KH PC ( )) (24)

4.7 Con lusion

At this stage, we end up with a model based on the following physi al and
hemi al prin iples:
 Mass balan e
 Ioni balan e
 AÆnity onstants
 Ideal gas law
 Henry's law
The more important hypothesis (with respe t to model reliability) is the
mass balan e hypothesis dedu ed from the rea tion s heme. This hypothesis
will therefore require to be validated in the sequel of the modelling approa h.
The mass balan e model an be used in this form for monitoring or
ontrol purpose. Indeed, using the approa hes developed in the framework
of systems with unknown inputs [10, 11, 12℄, the unknown rea tion rates an
be removed thanks to adequate state transformations [2℄.
 Mass Balan e Modelling of Biopro esses 789

Nevertheless, if the initial obje tive onsists in simulating the system,

then the rea tion rates ri (:) must be written with respe t to the state vari-
ables and to the system inputs (environmental variables). This step is mu h
more deli ate and a lot of hypotheses diÆ ult to verify are requested.

5 Modelling of the kineti s

5.1 Introdu tion

For some spe i purposes (optimal ontrol, simulation, predi tions, et .)
it is ne essary to have an analyti al expression relating the rea tion rates
to the state variables of the system. We have nevertheless to keep in mind
that these expressions are most of the time approximate relationships issued
from empiri al onsiderations. Therefore we leave the ba kground of physi al
modelling presented previously.
In this se tion we will see how to establish a hierar hy between the as-
sumed hypotheses in order to obtain a two reliability level des ription of the
kineti s.

5.2 The mathemati al onstraints

5.2.1 Positivity of the variables
A priori, some physi al onstraints that the model must respe t are known:
The variables must remain positive and they must be bounded if the amount
of matter entering in the biorea tor is bounded. These physi al onstraints
will impose onstraints on the stru ture of the ri (:). Some quantities (per-
entage, ratios, et .) must remain between known bounds. To guaranty that
the model respe ts this property, it should verify the following property:
Property 1 (H1) For ea h state variable i 2 [Li min ; Li max ℄, the eld _i
on the boundaries must be dire ted in the admissible spa e. In other words,
the following onditions must be satised:
i = Li min ) _i  0
i = Li max ) _i  0
Parti ular ase: We must have i = 0 ) _i  0. in order that variable
i remains positive.
790 O. Bernard

5.2.2 Variables that are ne essary for the rea tion

The se ond important onstraint whi h must be satised by the bio hemi al
kineti s is related to the rea tion s heme. A rea tion an not take pla e if
one of the rea tant ne essary for the rea tion is missing. This justies the
following property:
Property 2 If j is a rea tant of rea tion i, then j an be fa torised in ri :
ri (; u) = j ij (; u)
We verify then easily that j = 0 ) ri (; u) = 0
In the same way, for the rea tions asso iated to a biomass X , we have
the same property. Therefore a growth rea tion an be rewritten
ri (; u) = i (; u)X
The term i (; u) is alled the growth rate.

5.2.3 Example 1 ( ontinued)

Let us onsider the anaerobi digestion model given by equations (8) to (11)
and let us apply the state positivity prin iple:
X1 = 0 ) r1 (:)  0 (25)
X2 = 0 ) r2 (:)  0 (26)
S1 = 0 ) D(S1in S1 ) k1 r1 (:)  0 (27)
S2 = 0 ) D(S2in S2 ) + k2 r1 (:) k3 r2 (:)  0 (28)
Equations (25) and (26) are not very informative. In order that (27) and
(28) are respe ted whatever the experimental onditions, it requires:
r1 (:) = S1 1 (:) and r2 (:) = S2 2 (:)
Moreover, biomasses X1 and X2 are ne essary, respe tively for rea tions
1 and 2, and thus:
r1 (:) = 1 (:)X1 and r2 (:) = 2 (:)X2
Finally, we must have:
r1 (:) = S1 X1 1 (:) (29)
r2 (:) = S2 X1 2 (:) (30)
 Mass Balan e Modelling of Biopro esses 791

5.2.4 Phenomenologi al knowledge

We will exploit the available phenomenologi al knowledge (even if it is often
spe ulative) in order to propose an expression for the rea tion kineti s.
First, the laboratory experiments allows one to determine the variables
whi h a t on the rea tion rates. We have seen that the rea tant and some-
times the biomass must be found among these variables.
Then, we must know whether the rea tion is a tivated or inhibited by
these variables. It often happens that a variable is a tivating and that she
be omes inhibiting at high on entrations (toxi ity ee t).
Now, there remains to propose an analyti al expression whi h will take
into a ount the mathemati al onstraints so as the phenomenologi al knowl-
edge on the pro ess. For this, the modelling hoi es rely on one hand on
experimental observations (when they exist!) and on the other hand on
the available models in the literature. In all the ases, the parsimony prin-
iple will be privileged to guaranty that the models an be identied and
validated.
The following paragraph details the list of models that are often found
in the literature to des ribe some typi al rea tions. These examples are
indi ative and a very large number of dierent models an be found in the
literature, in parti ular to des ribe the growth rate [2, 1℄.

5.3 The growth rate

5.3.1 The Monod model
The most ommonly used model is the Monod [13℄ model whi h uses the
kineti s identied by Mi haelis-Menten for enzymati kineti s :
S
(S ) = max (31)
Ks + S
max is the maximal growth rate and Ks the half saturation onstant.
This simple model summarises the two main phases of the growth of a
mi roorganism:
 Unlimited growth, for high values of substrate (S >> KS ). The growth
rate is then onstant, equal to the maximal growth rate max
 The limited growth, for small values of substrate. In this ase the
growth rate is approximately proportional to the substrate.
792 O. Bernard

Note that the similitude between enzymati rea tion and growth of a
mi roorganism are often used to justify the analyti al expression of a rea tion
rate [14, 15℄.

5.3.2 Haldane model

The Haldane model, initially proposed for an enzymati rea tion an be used
to represent a substrate inhibiting the growth at high values [16℄:
S
 = max (32)
Ks + S + KS 2i
where Ki is an inhibition onstant. This model predi ts that the growth
rate is inversely proportional to the growth rate at high on entrations.

5.3.3 Multiple limitations

When two substrates S1 and S2 are simultaneously limiting the growth,
a usual way of modelling the rea tion rates is to take the produ t of two
Mi haelis-Menten kineti s:
 S1  S2 
 = max (33)
KS1 + S1 KS2 + S2
where KS1 and KS2 are the half saturation onstants asso iated respe tively
to substrates S1 and S2 .
If one of the substrate (say S1 ) is at high on entration, the growth rate
is then equivalent to a Monod model with respe t to the other substrate (.e.
S2 ).

5.4 Kineti s representation using neural networks

We expose brie y here an alternative method to represent the kineti s using
a neural network. The global model will then be omposed of a mass balan e
model based on O.D.E, and of a neural network for the rea tion rates. In
this sense it is an hybrid model. No a priori hypotheses are performed on the
kineti s, ex ept that we take into a ount some onstraints to guaranty that
the system traje tory keep an a eptable meaning. The kineti s represented
by the neural network are then dire tly identied along the training step.
 Mass Balan e Modelling of Biopro esses 793

ν1,1 ω1
S1
ω2
S2 ω3
µ
ωk

νm,1
ωn h
Sm
νm,n h

Input hidden output

layer layer layer

Figure 4: S heme of a neural network in luding a single hidden layer

Nevertheless, the variables whi h in uen e the kineti s must be determined.

These variables will onstitute the input of the neural network.
A s hemati view of the network is presented on Figure 4 for a single
hidden layer. The expression of the output of the network with respe t to
the inputs is as follows:

X
nh X
m
(S1 ; : : : ; Sm ) = !k ( kiSi ) (34)
k=1 i=1

where nh represents the number of neurons in the hidden layer. The

!k and ki are respe tively the weights of the input and outputs layers.
Fun tion  is the a tivating fun tion of the neuron. It is generally hosen
among a set of fun tions (sigmoides, hyperboli tangent, gaussian, et .).
The hoi e of the type of network and of the number of neurons is a rather
lassi al hoi e and we invite the reader to refer to [17℄ for more details.
On e the stru ture of the network has been hosen, the next step is the
training phase onsisting in identifying the networks weights. This opera-
tion is a bit spe i for hybrid systems and we refer to [18, 19℄ for more
explanations.
794 O. Bernard

6 Model validation

6.1 Introdu tion

The last modelling step is ertainly the most important, but it is also the
most often negle ted one. It is all the more important sin e we have seen
that it was ne essary to assume a great amount of spe ulative hypotheses.
Before using a model, it is important to validate it properly. This stage
follows generally the identi ation step whi h is not des ribed here.
The general obje tive for the validation is to verify that the model ts
the obje tives that have been xed. More pre isely, we will see how to
test separately the various hypotheses that have been assumed during the
model development:
 the rea tion s heme
 the qualitative model predi tions
 the model as a whole (rea tion s heme+kineti s+parameters)
It is important to note that the validation phase must be performed
from a data set whi h was not used to establish or to identify the model.
Moreover the new experiments that must be used to test the model validity
must signi atively dier from the previously used data set (otherwise it is
a test of the experimental reprodu ibility rather than a test of the model
validity). If these onditions are not respe ted, the model an not pretend
to be validated

6.2 Validation of the rea tion s heme

6.2.1 Mathemati al prin iple
The proposed pro edure relies on an important property, whi h is a on-
sequen e of the mass onservation within the biorea tor. As a result this
approa h will allow us to he k if the obtained mass balan e is onsistent
with the data.
Property 3 We assume that the nk matrix K has more rows than olumns
( n > k). This means that there are more variables than rea tions. In this
onditions, we have at least n k independent ve tors vi 2 Rn su h that:
vit K = 01k
 Mass Balan e Modelling of Biopro esses 795

By onvention, we normalise the rst omponent of the ve tor vi in order

to have vi 1 = 1
Consequen e : let us onsider the real variable wi = vit  , this variable
satises the following equation:
dwi
dt
= D(wi in wi ) vit Q( ) (35)
with wi in = vit in . Let us integrate (35) between two time instants t1 and
t2 . We rewrite this equation in order to let the omponents vij of ve tor vi
appear. It leads to:
X
n
vij j (t1 ; t2 ) = 1 (t1 ; t2 ) (36)
j =2
where
Z t2
j (t1 ; t2 ) = j (t2 ) j (t1 ) D( )(j in ( ) j ( )) Qj ( ( ))d
t1

The terms j (t1 ; t2 ) an be estimated from the experimental measurements

of j along time. An approximation of the integral an be omputed e.g.
using a trapeze formulae. Moreover if the sampling frequen y is not suÆ-
ient, the data will probably require to be interpolated. We re ommend for
this task to use spline fun tions whi h will at the same time smooth and
interpolate the data.
The relationship (36) is a linear relation linking the vij to the terms
j (t1 ; t2 ). Sin e the j (t1 ; t2 ) an be omputed between various time in-
stants t1 and t2 , (36) is a linear regression whose validity an be experimen-
tally tested.
Important remark: In fa t, relationship (36) is a linear regression
whi h will provide us with an estimate of the vij . These terms are related
with the oeÆ ients of the yield matrix K , and will in general allow to
estimate the value of these oeÆ ients.

6.2.2 Example 4
Let us onsider here the simple example of the growth of the lamentous
fungi Py noporus innabarinus (X ) on two substrates, glu ose ( arbon (C )
sour e) and ammonium (nitrogen (N ) sour e). We assume therefore that
the rea tion s heme is omposed by a single rea tion:
796 O. Bernard

N +C !X
The stoi hiometri matrix K asso iated to this rea tion is the following
( = (X N C )t ):
K = (1 k1 k2 )t ; and in = (0 Nin Cin )t (37)
Let us onsider the two following ve tors orthogonal to the olumns of K :
1
v = (1 0)t and v = (1 0 )t
1
1
k1 2
k2
We an then dene the following quantities:
Z t2
X (t1 ; t2 ) = X (t2 ) X (t1 ) + D( )X ( )
t1
Z t2
N (t1 ; t2 ) = N (t2 ) N (t1 ) D( )(Nin ( ) N ( ))d
t1
Z t2
C (t1 ; t2 ) = C (t2 ) C (t1 ) D( )(Cin ( ) C ( ))d
t1
whi h will allow us to rewrite the following regressions asso iated with v1
and v2 :
1
 (t ; t ) =  (t ; t ) (38)
X N
1 2
k1 1 2

X (t1 ; t2 ) =
1  (t ; t ) (39)
C 1 2
k2
It is now easy to verify if the relationships (38) and (39) are signi ative
from a statisti al point of view.
Figure (5) presents a validation example on the basis of a series of exper-
iment. The obtained regression is highly signi ative. This means that rela-
tions (38) and (39) are valid. As a onsequen e, the rows of matrix K , whi h
are orthogonal to v1 and v2 are ne essarily of the type K = (1 1 2 )t .
Therefore the rea tion s heme is valid, and subsequently the mass balan e
model as well.
Note that these te hniques lead also to the estimate of the yield oeÆ-
ients k1 and k2 .
 Mass Balan e Modelling of Biopro esses 797
2.5 2.5

A B
X estimated from N (g/l)

X estimated from S (g/l)

2 2

1.5 1.5

1 1

0.5 0.5
0.5 1 1.5 2 2.5 0.5 1 1.5 2 2.5
X measured (g/l) X measured (g/l)

Figure 5: Validation of the linear relationship relating X and N (A); X and C (B)

6.3 Qualitative model validation

For the third stage, we assume that the rea tion s heme, and therefore the
mass balan e model has been validated. We will then onsider a simulation
model onsisting of the mass balan e model plus the mathemati al expression
of the kineti s.
The rst think to do is to test whether the qualitative properties of the
model respe t the experimental observations.
The rst qualitative behaviour that we expe t the model to reprodu e
is the asymptoti behaviour obtained for onstant inputs. Will the model
predi t an equilibrium, or a more omplex behaviour (limit y le, haos,...)
in agreement with experiments ?
How do these properties evolve when the inputs vary ? For example, the
model will predi t that an equilibrium in a biorea tor is globally stable for
values of the dilution rate lower than a bound, and that for higher values
the equilibrium be omes unstable. Does it orrespond to the experimental
observations ?
More pre ise qualitative property on the type of transient allowed by
the model an also be ompared with experimental data. For some spe i
systems, these transients an be rather pre isely determined from a stru ture
analysis [20, 21, 22, 23℄.
Another qualitative riterion that an be dis ussed is the response of the
system at steady state to a hange in an input. Assume for example that an
in rease of input ui (whi h is then kept onstant) leads to a de rease in the
798 O. Bernard

steady state value of j : is it veried from an experimental point of view ?

6.3.1 Example
For example, Hansen and Hubbell (1980) study the ompetition between two
ba terial spe ies in a hemostat. The rea tion s heme is omposed of two
growth rea tions:
k1 S !X 1

k2 S !X 2

The growth rate asso iated to these rea tions is assumed to be of Monod
type, i.e.:
S
i (S ) = max i
S + Ks i
where max i and Ks i are the maximum growth rate and the half saturation
onstant asso iated with substrate S for spe ies i.
Hansen and Hubbell showed that the winner of the ompetition predi ted
by the model depends on the dilution rate. More pre isely, the winner is
the spe ies with the smaller ratio Ji = max Ks i . The omparison of the 2
i D
ratios J1 and J2 leads to the study of the quantity r = max 1 max 2 with
max2 D
respe t to the threshold value KKss 12 1. If we assume that we are in the
ase where D < max 1 < max 2 , then spe ies 2 wins for a dilution rate
lower than D0 = max 2 KKss 11 Kmax
s2
1 Ks 2 , whereas for higher values, it is spe ies
1 (see gure 6). These qualitative properties are veried experimentally (see
Figure 7).

6.4 Global model validation

This is the lassi al way of validating a model: the simulation results are
quantitatively ompared to experimental data. The most popular riterion
is the least square riterion whi h is omputed as follows for a data set of N
measurements:
XN
J = j^(ti )  (ti )j2
i
where ^(ti) is the simulated value of the state  at the sampling instant ti .
The riterion an be improved by weighting ea h omponent of the state
j by a oeÆ ient whi h takes into a ount the mean value of j and the
measurement a ura y for this variable.
 Mass Balan e Modelling of Biopro esses 799

µ max 1 − µ max 2

µ max 2 − D

K s1
−1
Ks2

µ max 1
−1
µ max 2
Species 2 survives Species 1 survives Both species disappear

0 µ max 2 µ (Sin) µ max 1 D

D0 1

Figure 6: Competition in a hemostat with respe t to the dilution rate (dis ussion of
the quantity max 1 max 2 with respe t to ks 1
max 2 D ks 2 1). We onsider here the ase where
D < max 2 < max 1

This riterion should be minimum. In theory, the residuals (i.e. ^  )

must be studied from a statisti al point of view. In the ideal ase, it should
have properties omparable to those of the measurement noise: it should
at least be zero on average, and more pre isely one an expe t a gaussian
distribution [25℄.
In this approa h, the model is onsidered as a whole. If the residual
analysis is not good, in the ase where the previous validation steps (rea tion
s heme and qualitative riteria) have not been performed properly it would
be impossible to know the ause of the problem. This riterion does not give
any lue on the stru tural validity of the model (underlying rea tion s heme,
qualitative properties), on the validity of the type of rea tion rate modelling
used or on the orre tness of parameter values.
If the two rst validation steps have been su essfully fullled, the prob-
lem is probably due to a an errati parameter estimation.
In pra ti e, in the framework of biote hnologi al systems, as it is diÆ ult
to validate stri to sensu these models, one will be satised with a good
800 O. Bernard

Figure 7: Experimental validation of the qualitative model behaviour. Quantitative

model predi tions are represented as well. The qualitative model predi tions are veried
for: a) Two spe ies (Es heri hia oli, strain C-8 and Pseudomonas aeruginosa, strain
PA0283 whi h dier from their half-saturation onstants. b) Two strains of Es heri hia
oli whi h dier from their maximal growth rates. d) Coexisten e obtained with 2 strains
of Es heri hia oli whi h have the same parameter Ji . Figure ) represents the ee t of
nalidixi a id on the maximal growth rate for the 2 onsidered strains C-8. (from [24℄)

visual adequation between simulations and data. This subje tive riterion
an be reinfor ed by an analysis of the orrelation between predi tions and
measurements.

6.4.1 Example
The following validation example presents the results obtained with the
anaerobi digestion model exposed throughout the paper. Figures 8 and
9 present model simulations ompared to dire t measurements [9℄. The pe-
riods of time onsidered for the alibration step are shown on the gures.
The model orre tly reprodu es the behaviour of the system for the on-
sidered period in spite of the fa t that it has been alibrated only using
 Mass Balan e Modelling of Biopro esses 801

steady state measurements.

Indeed Figure 8 shows that the ontinuously measured variables (i.e.
gaseous ow rate and pH) are well predi ted. It is worth noting that these
simulations also orre tly reprodu e the ee t of the disturban es indu ed by
pump failures (around day 45). Remark also that the pH predi tions mat h
quite well the dire t measurements although pH measurements have not
been used to alibrate the model parameters. However the model predi ts
a more severe pH drop during the destabilisation phase (days 21-25). This
may be due to an underestimation of the buer apa ity (i.e. the alkalinity
of the system). It an be noti ed that during the destabilisation period the
gases are underestimated by the model.
The model simulations are also in good agreement with the o-line data
(Figure 9). Even if S1 is a variable that stands for the various omponents
of the COD that an be rather dierent along the experiment, the adequa y
between model and measurements is good. The rea tion of the model to the
overloading produ ed on day 68 seems to be slower than the pro ess, so that
the a umulation starts less rapidly in the model.
The main quality of the model is its ability to predi t the destabilisation
of the plant. This was not obvious sin e only equilibrium data have been
used for the model alibration and the data obtained during the destabilisa-
tion phases were not used. The quality of the model justies its integration
in an on-line monitoring pro edure in order to early dete t a possible desta-
bilisation [26℄. The model is also used to derive a robust ontrol algorithm,
that is insensitive to the main modelling un ertainties and that avoid the
plant destabilisation [27℄.

7 Mass balan e models properties

7.1 Boundness and positivity of the variables

We have seen in paragraph 5.2.1 that the models must be designed in order
to meet onstraints like the positivity of the state variables.
We will see here that the models based on mass balan es are of the type
BIBS (bounded input bounded state). To show this property, we use the
following hypotheses whi h are veried for the mass balan e based systems:
Hypothesis 1 (H2) There exists a ve tor v+ whose omponents are stri tly
802 O. Bernard

150 250

200

Qco2 (l/h)

Qch4 (l/h)
100
150

100
50

50

0 0
10 20 30 40 50 60 70 10 20 30 40 50 60 70
Time (days) Time (days)
350 7.5

300 7

250
Qtot (l/h)

6.5
200

pH
6
150
5.5
100

50 5

0 4.5
10 20 30 40 50 60 70 10 20 30 40 50 60 70
Time (days) Time (days)

Figure 8: Comparison between simulation results and measurements for the gaseous ow
rates and the pH. The periods onsidered for the alibration step are represented on the
time axis

positive, su h that:
v+ K = 01k
Consequen e: Let us onsider the s alar quantity w+ = v+  . It veries
the following equation: (35):
dw+
dt
= D(win+ w+ ) v+ Q( ) (40)
We have to assume an hypothesis for Q( ), whi h is veried in most of
the ases:
Hypothesis 2 (H3) There exists a positive real a and a real b, su h that
Q( ) an be ompared to a linear expression as follows:
v+ Q( )  av+  + b
 Mass Balan e Modelling of Biopro esses 803
10 40

35
8

VFA (mmol/l)
30
COD g/l 25
6
20
4 15

10
2
5

0 0
10 20 30 40 50 60 70 10 20 30 40 50 60 70
Time (days) Time (days)
80 70

70 60
60

C (mmol/l)
Z (mmol/l)

50
50
40
40
30
30
20
20

10 10

0 0
10 20 30 40 50 60 70 10 20 30 40 50 60 70
Time (days) Time (days)

Figure 9: Comparison between simulation results and measurements for COD, VFA,
alkalinity and total inorgani arbon. The periods onsidered for the alibration step are
underlined

This hypothesis is veried if v+ Q( ) = 0, or if Q( ) is des ribed by

Henry's law (see se tion 4.4).
Property 4 If hypotheses (H1), (H2) and (H3) are veried, then the system
is BIBS.

Proof: The dynami s of w+ an be bounded as follows:

dw+
 (D + a)( Dwin b w )
+

dt D+a
+
(41)
+
if we apply propertyP1, we an dedu e: w+  max(w+ (0); DwDin+a b ).
In other words, wi+ i is bounded. Sin e wi+ > 0, the state variables i
are bounded.
804 O. Bernard

7.2 Equilibrium point and lo al behaviour

7.2.1 Introdu tion
In this se tion we brie y re all the prin iples of the studies of the model
properties. We invite the reader to onsult [28℄ for more details.
Generally, the biorea tor models are non linear (e.g. they often have
multiple steady state), and they are of high dimension (large number of state
variables). They often have a large number of parameters, whi h often in-
tervene in nonlinear fun tions (nonlinearity with respe t to the parameters).
Nevertheless, for dimensions greater than 3, it be omes very diÆ ult to
hara terise the behaviour of a dynami al system. We will however show
that the mass balan e based model have stru tural properties whi h make
easier the system understanding.
In this paragraph, we onsider a general dynami al system:
d
dt
= f (; u) (42)
We keep in mind that f (; u) = Kr( ) + D(in  ) Q( ). We will
onsider here the ase where u = (D; in ) is onstant.

7.2.2 Equilibrium points and lo al stability

The equilibrium points are obtained for ddt = 0 when the inputs are main-
tained onstant.
The non linear systems generi ally dier from linear systems sin e they
an have multiple equilibrium points.
The rst step in the model analysis onsists in testing if these equilibrium
points are lo ally stable. We onsider the ja obian matrix of the linearised:
Df
J ( ) = ( )
D
The equilibrium 0 is lo ally stable if and only if all the eigenvalues of
J (0 ) have a negative real part. If there exists an eigenvalue with positive
real part, the equilibrium is unstable. We an not on lude on the system
stability if none eigenvalues have a positive real part but one (at least)
eigenvalue has a zero real part.
 Mass Balan e Modelling of Biopro esses 805

7.2.3 Global behaviour

The dynami s of a nonlinear system an be very ompli ated, and om-
plex behaviours like limit y les, haos, et . an appear in addition to the
equilibria. It is therefore important to test whether a unique lo ally stable
equilibrium is globally stable. In other words if for any initial onditions the
traje tories will onverge toward this equilibrium.
The standard method to prove that an equilibrium is globally stable relies
on the Lyapunov [28℄ approa h. However it is often diÆ ult to nd a Lya-
punov fun tion for a biologi al system. One an refer to [29℄ for onstru tive
methods to nd Lyapunov fun tions in a large lass of growth models.

7.2.4 Asymptoti behaviour

We have seen in paragraph 6.2.1 that in the general ase where n > k, there
exists n k ve tors vi in the kernel of K T . These ve tors allow to ompute
the quantities wi = vit  whose dynami s satises equation (35).
Moreover, there are often q ve tors vi0 among the vi whi h verify:
vi0 t Q( ) = 0 (43)
The dynami s of the asso iated wi0 is then very simple:
dwi0
dt
= D(wi0in wi0 ) (44)
In the onditions that we onsider (i.e. onstant D and in ), the solutions
of (44) asymptoti ally onverge towards wi0in . This means that the solutions
of system (42) will onverge towards the hyperplane vi0 t  = 0.
The state of the system will then asymptoti ally onverge toward the
ve torial subspa e of dimension n q, whi h is orthogonal to the q ve tors
vi0 . This allows to simplify the study of the n dimensional system (42) into
a n q dimensional system.

7.2.5 Example 4 ( ontinued)

Let us onsider the model of fungal growth (equation 37). We will moreover
assume that the kineti s has been represented by a Monod law with respe t
to the 2 substrates C and N :
C N
r( ) = max X (45)
KC + C KN + N
806 O. Bernard

The two ve tors v1 and v2 identied in paragraph (6.2.2) verify straightfor-

wardly equation (43).
Therefore when t ! +1, X + kN1 ! Nkin1 and X + kC2 ! Ckin2 .
The study of the 3 dimensional system is then simplied into the study
of the following system in dimension 1:
dX
dt
= max K C+inC k2 Xk X K N+inN k1 Xk X X DX (46)
C in 2 N in 1

One will verify that this system has three real equilibrium points (one of
them being the trivial equilibrium X = 0). These equilibria, in in reasing
order, are respe tively lo ally stable, unstable and lo ally unstable. With
respe t to the parameters values, the equilibria will be positive (and therefore
admissible) or not. For the parametri domains where there exists a single
positive equilibrium, this equilibrium is globally stable.

8 Con lusion

We have presented a onstru tive and systemati method to develop bio-

pro ess models in 4 steps. Let us re all that the modelling of a biopro ess
must be performed in the framework of a learly identied obje tive. The
modelling must orrespond to the quality and the quantity of the available
information so that the model an be orre tly validated and identied.
The rst modelling steps onsists in gathering the physi al and hemi al
prin iples that an apply to the system and to assume a rea tion s heme in
order to obtain the mass balan e model.
In a se ond step, one must take benet of the onstraints that the model
must verify and use the empiri al relationships to nd an analyti al expres-
sion for the rea tion kineti s.
The third step onsists in identify the model parameters by separating
those who are related to the mass balan es (yield oeÆ ients), those who
are related with the used physi al prin iples (aÆnity onstants, transfer
onstants, et .) and those who intervene in the rea tion rates.
Finally, the ultimate modelling step must not be negle ted: namely the
model validation. During this last step the model quality must be tested
using the more obje tive as possible riteria. The validity of the model must
be assessed along its ability to properly represent the mass balan e, to repro-
du e orre tly the qualitative features of the data, and to t quantitatively
 Mass Balan e Modelling of Biopro esses 807

the data. The important point is that the data whi h must be used for
model validation must not have been already used in the model onstru tion
phase. During the validation step, not only the quality of the model will
be assessed, but also its validity domains: the working domains (in terms of
state variable and inputs) where the model is satisfa tory.
To on lude, we insist on the fa t that the modelling step an be very
long and expensive, but the quality of a model is a ne essary onditions to
ensure that a ontroller or an observer based on it will properly work.

Appendix A. Theoreti al determination of the dimension of K

Let us integrate equation (7) between 2 time instants t and t + T :
Z t+T Z t+T
 (t + T )  (t) D(in( )  ( ))) + Q( ( ))d = K r( ( ))d ;
t t
(47)
Let us denote:
Z t+T
v(t) =  (t + T )  (t) D(in ( )  ( ))) + Q( ( ))d
t
and Z t+T
w (t ) = r( ( ))d
t
Equation (47) an then be rephrased:
v(t) = K w(t) (48)
The ve tor v(t) an be estimated along time on the basis of the available
measurements. The integral value an be estimated e.g. with a trapeze
approximation.
To avoid onditioning problem and to give the same weighting to all the
state variables, we normalise the data ve tors u(ti ) as follows:
v(t ) e(v)
v~(ti ) = pi
N(v)
where e(v) is the average value of v(ti ), and (v) their standard deviation.
Now the question of the dimension of matrix K an be formulated as
follows: what is the dimension of the image of K , in other words, what is
808 O. Bernard

the dimension of the spa e where u(t) lives. Note that we are looking for
a full rank matrix K . Otherwise, it would mean that the same dynami al
behaviour ould be obtained with a matrix K of lower dimension.
Determining the dimension of the v(t) spa e is a lassi al problem in
statisti al analysis. It orresponds to the prin ipal omponent analysis that
determines the dimension of the ve torial spa e spanned by the ve tors ki ,
rows of K . To rea h this obje tive, we onsider matrix U obtained from a
set of N re ording of v(t):
V = (~v (t1 ); : : : ; v~(tN ))
We will also onsider the asso iated matrix of rea tion rates, whi h is
unknown:
W = (w(t1 ); : : : ; w(tN ))
We assume that matrix W is of full rank. This means rst that there are
more measurements than rea tions. It means also that the rea tions are in-
dependent (none of the rea tion rates an be written as a linear ombination
of the other ones).
Property 5 For a matrix K of rank k, if W has full rank, then the n 
n matrix M = V V T = KW W T K T has rank k. Sin e it it is a positive
symmetri matrix, it an be written, by:

M = P t P
where P is an orthogonal matrix (P T P = I ) and
0 1
1 0 ::: 0
B
B C
C
B
B
0 2 0 0C
C
B
B ..
.
... C
C
B C
=B
B
B k C
C
C
B
B 0 C
C
B
B C
B . . . .. C
 . C
A
0 ::: 0
with i 1  i > 0 for i 2 f2; :::; kg.
 Mass Balan e Modelling of Biopro esses 809

Proof: it is dire t appli ation of the singular de omposition theorem [30℄.

Sin e rank (M ) = rank () = k, it provides the result.
Now from a theoreti al point of view it is possible to determine the
number of rea tions in the rea tion s heme: it orresponds to the rank of K
or, in other words, to the number of non zero singular values of V V T .
In the reality, the noises due to model approximations, measurement
errors or interpolation perturb the analysis. Therefore in pra ti e there are
no zero eigenvalues for the matrix M = V T V .
The question is then to determine the number of eigenve tors that must
be taken into a ount in order to represent a reasonable approximation of
the data v(t). To solve this problem, let us remark that the eigenvalues i of
M orrespond to the varian e asso iated with the orresponding eigenve tor
(inertia axis).
The method will then onsist in sele ting the p rst prin ipal axis whi h
represent a total varian e larger than a xed threshold.
810 O. Bernard

Referen es

[1℄ J. E. Bailey and D. F. Ollis, Bio hemi al engineering fundamentals.

M Graw-Hill, 1986.
[2℄ G. Bastin and D. Do hain, On-line estimation and adaptive ontrol of
biorea tors. Amsterdam: Elsevier, 1990.

[3℄ F. Mosey, \Mathemati al modelling of the anaerobi digestion pro ess:

regulatory me hanisms for the formation of short- hain volatile a ids
from glu ose," Water S ien e and Te hnology, vol. 15, pp. 209{232,
1983.
[4℄ D. Hill and C. Barth, \ A dynami model for simulation of animal waste
digestion," Journal of the Water Pollution Control Asso iation, vol. 10,
pp. 2129{2143, 1977.
[5℄ R. Moletta, D. Verrier, and G. Albagna , \Dynami modelling of anaer-
obi digestion," Wat.Res., vol. 20, pp. 427{434, 1986.
[6℄ D. Costello, P. Greeneld, and P. Lee, \ Dynami modelling of a single-
stage high-rate anaerobi rea tor - I. Model derivation," Water Re-
sear h, vol. 25, pp. 847{858, 1991.

[7℄ D. Batstone, J. Keller, B. Newell, and M. Newland, \Model development

and full s ale validation for anaerobi treatment of protein and fat based
wastewater," Water S ien e and Te hnology, vol. 36, pp. 423{431, 1997.
[8℄ J. Mer huk, \Further onsiderations on the enhan ement fa tor for oxy-
gen absorption into fermentation broth," Biote hnol. & Bioeng., vol. 19,
pp. 1885{1889, 1977.
[9℄ O. Bernard, Z. Hadj-Sadok, D. Do hain, A. Genovesi, and J. Steyer,
\ Dynami al model development and parameter identi ation for an
anaerobi wastewater treatment pro ess," Biote h.Bioeng., no. 75,
pp. 424{438, 2001.
[10℄ P. Kudva, N. Viswanadham, and A. Ramakrishna, \Observers for linear
systems with unknown inputs," IEEE Trans. Autom. Contr., vol. AC-
25, no. 1, pp. 113{115, 1980.
 Mass Balan e Modelling of Biopro esses 811

[11℄ M. Hou and P. Mller, \Design of observers for linear systems with un-
known inputs," IEEE Trans. Autom. Contr., vol. AC-37, no. 6, pp. 871{
875, 1991.
[12℄ M. Daroua h, \On the novel approa h to the design of the unknown
input observers," IEEE Trans. Autom. Cont., vol. 39, no. 3, pp. 698{
699, 1994.
[13℄ J. Monod, Re her hes sur la roissan e des ultures ba triennes. Paris,
Fran e: Hermes, 1942.
[14℄ L. A. Segel, Modeling Dynami Phenomena in Mole ular and Cellular
Biology. Cambridge: Cambridge University Press, 1984.

[15℄ L. Edelstein, Mathemati al Models in Biology. New York: Random

House, 1988.
[16℄ J. Andrews, \A mathemati al model for the ontinuous ulture of mi-
roorganisms utilizing inhibitory substrate," Biote hnol. & Bioeng.,
vol. 10, pp. 707{723, 1968.
[17℄ J. Hertz, A. Krogh, and R. Palmer, Introdu tion to the Theory of Neural
Computation. Addison-Wesley, 1991.

[18℄ L. Chen, O. Bernard, G. Bastin, and P. Angelov, \Hybrid modelling of

biote hnologi al pro esses using neural networks," Contr. Eng. Pr ati e,
vol. 8, pp. 821{827, 2000.
[19℄ A. Karama, O. Bernard, A. Genovesi, D. Do hain, A. Benhammou,
and J.-P. Steyer, \Hybrid modelling of anaerobi wastewater treatment
pro esses," Wat. S i. Te hnol., vol. 43, no. 1, pp. 43{50, 2001.
[20℄ C. Jeries, \Qualitative stability of ertain nonlinear systems," Linear
Algebra and its Appli ations, vol. 75, pp. 133{144, 1986.

[21℄ E. Sa ks, \A dynami systems perspe tive on qualitative simulation,"

Artif. Intell., vol. 42, pp. 349{362, 1990.

[22℄ O. Bernard and J.-L. Gouze, \Transient behavior of biologi al loop mod-
els, with appli ation to the Droop model," Mathemati al Bios ien es,
vol. 127, no. 1, pp. 19{43, 1995.
812 O. Bernard

[23℄ J.-L. Gouze, \Positive and negative ir uits in dynami al systems,"

Journal Biol. Syst., vol. 6, no. 1, pp. 11{15, 1998.
[24℄ S. R. Hansen and S. P. Hubbell, \Single-nutrient mi robial ompeti-
tion," S ien e, vol. 207, no. 28, pp. 1491{1493, 1980.
[25℄ E. Walter and L. Pronzato, Identi ation de modeles parametriques.
Masson, 1994.
[26℄ O. Bernard, Z. Hadj-Sadok, and D. Do hain, \Dynami al modelling and
state estimation of anaerobi wastewater treatment plants," in Pro eed-
ings of ECC99 (CDROM), Karlsruhe, Germany, 1999.
[27℄ O. Bernard, M.Polit, Z. Hadj-Sadok, M.Pengov, D. Do hain,
M.Estaben, and P.Labat, \Advan ed monitoring and ontrol of anaer-
obi wastewater treatment plants: software sensors and ontrollers for
an anaerobi digestor," Wat. S i. Te hnol, vol. 43, no. 7, pp. 175{182,
2001.
[28℄ H. Khalil, Nonlinear Systems. Ma millan Publishing Company, 1996.
[29℄ B. Li, \Global asymptoti behavior of the hemostat: General response
fun tions and dierent removal rates," SIAM Journal, vol. 59, 1998.
[30℄ R. Horn and C. Johson, Matrix analysis. Cambridge Univesity Press,
1992.