foods

Article
Parameter Estimation of Dynamic Beer Fermentation Models
Jesús Miguel Zamudio Lara 1,2 , Laurent Dewasme 1 , Héctor Hernández Escoto 2
and Alain Vande Wouwer 1, *

1 Systèmes, Estimation, Commande et Optimisation, Université de Mons, 7000 Mons, Belgium

2 Departamento de Ingeniería Química, Universidad de Guanajuato, Guanajuato 36050, Mexico
* Correspondence: alain.vandewouwer@umons.ac.be

Abstract: In this study, two dynamic models of beer fermentation are proposed, and their parameters
are estimated using experimental data collected during several batch experiments initiated with
different sugar concentrations. Biomass, sugar, ethanol, and vicinal diketone concentrations are
measured off-line with an analytical system while two on-line immersed probes deliver temperature,
ethanol concentration, and carbon dioxide exhaust rate measurements. Before proceeding to the
estimation of the unknown model parameters, a structural identifiability analysis is carried out to
investigate the measurement configuration and the kinetic model structure. The model predictive
capability is investigated in cross-validation, in view of opening up new perspectives for monitoring
and control purposes. For instance, the dynamic model could be used as a predictor in receding-
horizon observers and controllers.

Keywords: parameter estimation; mathematical modeling; beer fermentation; food industry

Citation: Zamudio Lara, J.M.; 1. Introduction

Dewasme, L.; Hernández Escoto, H.; Beer is the most consumed alcoholic beverage worldwide and is produced by the
Vande Wouwer, A. Parameter fermentation of sugars in the wort by yeasts [1]. The production of beer in 2019 was 1912
Estimation of Dynamic Beer
million hectoliters (hl), while in 2020, the production reduced to 1820 million hl due to
Fermentation Models. Foods 2022, 11,
the COVID-19 pandemic [2]. Beer production is a complex biochemical process in which
3602. https://doi.org/10.3390/
the main ingredients are water, malt (sugar source), yeast, and hops [3]; however, other
foods11223602
products can be added, such as fruits, chocolate, and coffee grains, among others.
Academic Editors: Jose A. Egea, The fermentation stage is crucial to guarantee good quality beer since it is when all
Carlos Vilas and Míriam R. García the nutrients, flavor, and odor components are produced, in addition to ethanol. At this
stage, yeast is introduced in the wort (broth that is rich in sugars) from the boiling stage
Received: 3 October 2022
Accepted: 7 November 2022
at the desired temperature. The main chemical reaction is the conversion of these sugars
Published: 11 November 2022
into ethanol and carbon dioxide, along with biomass growth and heat generation. At the
same time, several secondary reactions occur, generating several components at lower
Publisher’s Note: MDPI stays neutral concentrations that contribute to the flavor and aroma characteristics.
with regard to jurisdictional claims in
To enhance the fermentation, several factors, such as yeast pitching rate, dissolved
published maps and institutional affil-
oxygen, batch pressure, and temperature, must be taken care of by the brewers [4]. Among
iations.
these factors, temperature is important as it helps accelerate the fermentation but needs
to remain within controlled bonds to avoid yeast death (above 30 °C), the production of
undesirable byproducts, and the growth of bacteria, damaging the final product. Therefore,
Copyright: © 2022 by the authors.
rigorous control of the temperature inside the fermenter must be exercised to ensure
Licensee MDPI, Basel, Switzerland.
product quality and alleviate variations between batches.
This article is an open access article In the brewing industry, time-varying temperature profiles are established along the
distributed under the terms and fermentation process in order to alleviate the above-mentioned potential issues [5]. Looking
conditions of the Creative Commons for the appropriate temperature profile is, however, not an easy task, and experimental
Attribution (CC BY) license (https:// determination can be time-consuming. Model-based optimization is, therefore, an ap-
creativecommons.org/licenses/by/ pealing alternative. Dynamic models can be useful not only to optimize the operating
4.0/).

Foods 2022, 11, 3602. https://doi.org/10.3390/foods11223602 https://www.mdpi.com/journal/foods

Foods 2022, 11, 3602 2 of 20

conditions, but also to design state estimators reconstructing online non-measured variables
or designing controllers to ensure close setpoint tracking [6].
To this end, Gee and Ramirez [7] proposed a detailed model of beer fermentation
describing biomass growth and the production of flavor compounds through macroscopic
reactions inferred from biological pathways. This model segregates the sugars into glucose,
maltose, and maltotriose. The derivation of the corresponding sugar uptake kinetics is,
therefore, the center of interest, and the related parameters are assumed to be temperature-
dependent. However, this model is unlikely to be applicable for control purposes since
it involves variables that require specific monitoring equipment beyond the standards
of most fermenters. Andrés-Toro et al. [8], conversely, proposed to segregate the yeast
(biomass) into three types: lag, active, and death, while the sugars are considered as a
whole. Sugar monitoring, therefore, appears simpler, and the model also takes account
of industrial operational characteristics as well as the undesired beer flavor caused by
ethanol and byproduct (diacetyl and ethyl acetate) formation. The main drawback of this
model is the difficulty in obtaining accurate biomass data. Trelea et al. [9] developed
what can be considered as, so far, the most practical model, where only three states are
considered: the dissolved carbon dioxide concentration used as an image of the growing
biomass, the ethanol concentration, and the sugar concentration. The main advantage of
this latter model lies in its practical control-oriented description of the fermentation process,
considering variables that can be easily measured and tracked.
The objective of this work is to revisit these classical models, propose a few adap-
tations, and develop a thorough study of the parameter estimation problem based on
a popular fermentation device, e.g., a 30-L stainless-steel Grainfather® fermenter. Two
alternative mathematical models are considered, one based on yeast (biomass) and the
other on carbon dioxide. The difference between these models is discussed in terms of
biological interpretation, bioreactor instrumentation, and data collection (i.e., parameter
estimation, model validation, and process control perspectives). As a result, models with
good predictive capability are proposed together with their experimental validation.
This paper is organized as follows. The next section describes the experimental setup,
while Section 3 presents a review of dynamic models of beer fermentation, together with
possible model adaptations required to represent the considered case study. Section 4
develops a structural identifiability analysis based on the software tool Strike Goldd [10].
Section 5 introduces a parameter identification procedure, including parametric sensitivity
analysis and model validation. The last section draws the main conclusions of this work
and discusses the monitoring and control perspectives.

2. Beer Fermentation Experimental Set-Up

The pilot plant consists of a stainless-steel conical fermenter (30 L, Grainfather® ), which
has a built-in sensor to measure the temperature of the liquid content. This sensor is paired
to a control system connected to a glycol chiller (Grainfather® ) to keep the temperature
regulated. Ethanol and carbon dioxide concentrations are measured online, respectively,
with a tilt® hydrometer and a Plaato® airlock.
The hydrometer is introduced in the wort and keeps floating in a tilted position, mea-
suring the specific gravity which also allows, based on some predetermined correlations,
for assessing the percentage of alcohol. The sensor also has an integrated temperature
sensor. The airlock consists of four components: a lid, a bubbler, a Tritan, and a smart part
(containing the temperature and infrared sensors). This device measures the evenly-sized
bubbles of carbon dioxide released by the wort and converts them into liters of CO2 . The
data are stored and displayed in the Brewblox® interface.
Besides the two online probes, a CDR BeerLab® analyzer is used to obtain offline
measurements of sugar, ethanol, and vicinal diketone (VDK) concentrations. The scheme
of the full experimental setup is displayed in Figure 1.
Foods 2022, 11, 3602 3 of 20

Figure 1. Stainless-steel fermenter pilot plant monitoring set-up.

In this study, ale beer fermentation was considered and carried out at a temperature
ranging from 17 to 26 °C. To obtain the wort, the malt is crushed using a mill. Indeed, the
grain must only be broken, but not grounded. The next step is mashing, where sugars
are obtained from starch. The crushed grain is added to a boiler tank (35 L, Grainfather® )
containing 19 L of water at 48 °C. The mashing step consists of four stages at different
temperatures and times described in Table 1. Once mashing is finished, the grain is rinsed
with water at 75 °C until 24 L of wort is obtained. Usually, the quantity of added water is
8 L. Then, the wort is boiled to sterilize the liquid. The latter step is carried out for 80 min
at 100 °C. Hop is added at 40 min and 65 min. Eventually, the wort is cooled down as fast
as possible to the desired fermentation temperature with the help of a counter-flow wort
cooler. The cold wort is transferred in the fermentation tank, filled up to 17 L.
A set of four isothermal batch fermentations without agitation are carried out using
different operating conditions described in Table 2. Each experiment is carried out once,
but two replicates are taken and analyzed for each sample. The total sampling volume
represents less than 10% of the initial wort volume (17 L), a condition to neglect the volume
changes. Samples are taken every 2 to 3 h during the first 36 h. After this period, the
process enters a stationary phase and the sampling time is therefore adapted at irregular,
longer, time intervals. To analyze the samples with the CDR BeerLab® , it is necessary to
achieve preprocessing, including degasification and centrifugation to eliminate everything
that could interfere during the measurement.

Table 1. Operating conditions of the scheduled mash steps: temperatures and times.

Step Temperature (°C) Time (min)

1 48 10
2 62 30
3 71 20
4 78 5
Foods 2022, 11, 3602 4 of 20

Table 2. Operating conditions of each fermentation batch: duration, temperature, and initial sugar
and biomass concentrations.

Temperature
Experiment Time (h) S0 (g/L) X0 (g/L)
(°C)
1 72 19 88 7.05 × 10−1
2 54 19 72 5.29 × 10−1
3 96 21 75 7.05 × 10−1
4 72 28 81 7.05 × 10−1

3. Mathematical Models of Beer Fermentation

Beer fermentation has been studied extensively, providing various mathematical
models, and this section proposes a brief description of three of them, considered as
milestones in the research field and used as a reference in the upcoming model development.
These mathematical models can provide precious support to design the fermentation
operating conditions.

3.1. The Model of Gee and Ramirez

This model published in 1994 includes the main components of fermentation, e.g., sug-
ars, biomass, and ethanol, as well as amino acids, fusel alcohols, VDKs, and acetaldehydes,
which impact the flavor, often in an undesirable way [7]. Esters also have an important role
in the aroma and may add some pleasant character in moderate ranges, but undesired hard
fruity tastes at higher levels.
The sugar uptake model reads as follows:

dG
= −µG X, (1a)
dt
dM
= −µ M X, (1b)
dt
dN
= −µ N X, (1c)
dt
where G, M, and N, respectively, stand for glucose, maltose, and maltotriose. The specific
growth rates are built upon classical kinetic activation (Monod law) and inhibition factors
and are given by:

µG G
µG = , (2a)
KG + G
0
KG
µM M
µM = 0 , (2b)
K M + M KG + G
0
KG K 0M
µN N
µN = 0 + G K0 + M , (2c)
K N + N KG M

where µi are maximum rate constants (i = G, M, N), while Ki and K 0j (j = G, M) are,

respectively, the half-saturation and inhibition constants, all assumed to depend on the
temperature following an Arrhenius law of the form:
 
B
r = Aexp , (3)
RT

where A and B are, respectively, the Arrhenius frequency factor and the activation energy.
R is the ideal gas constant.
Biomass production is represented by the mass-balance ODE:

dX
= µ X X, (4)
dt
Foods 2022, 11, 3602 5 of 20

where
µ x = YXG µ1 + YXM µ2 + YXN µ3 (5)
is the biomass growth rate, a function of the several sugar intake rates. YXi (i = G, M, N)
are the yield coefficients of biomass with respect to the specific sugars.
The ethanol concentration is assumed to evolve proportionally to the variations of the
sugar concentrations, resulting in an algebraic equation of the form:

E = E0 + YEG ( G0 − G ) + YEM ( M0 − M ) + YEN ( N0 − N ), (6)

where YEi (i = G, M, N) are the yield coefficients of ethanol with respect to the con-
sumed sugars.
Three main amino acids are considered, which are responsible for the formation of
flavor compounds in the beer like the fusel alcohols. The amino acids are described by the
following differential equations:

dξ ξ dX ξ
= −Yξ = −Yξ µ X X, (7)
dt Kξ + ξ dt Kξ + ξ

where Yξ and Kξ are, respectively, the yield coefficients and inhibition constants with
species ξ = leucine (L), isoleucine (I), and valine (V).
The impact of fusel alcohols is a plastic, solvent-like flavor. Moreover, some exper-
iments achieved in [11] have also linked higher alcohol levels with physiological effects
associated with hangovers. The four fusel alcohols represented in the model are isobutyl
alcohol (IB), isoamyl alcohol (IA), 2-methyl-1-butanol (MB), and n-propanol (P).

dIB
= YIB µV X, (8a)
dt
dI A
= YI A µ L X, (8b)
dt
dMB
= YMB µ I X, (8c)
dt
dP
= YP (µV + µ I ) X, (8d)
dt
where Yζ (ζ = IB, I A, MB, P) are yield coefficients, and µξ are specific rates expressed as
µξ = − X1 dξ ξ
dt = Yξ µ X Kξ +ξ ( ξ = L, I, V ).
Esters contribute mainly to the aroma of the beer due to their high volatility. In
moderate concentrations, they can confer a pleasant character to the beer. However, once in
excess, the aroma becomes overly fruity, which is undesired by most consumers. Principal
esters are ethyl acetate (EA), ethyl caproate (EC), and isoamyl acetate (I Ac).

dEA
= YEA (µG + µ M + µ N ) X, (9a)
dt
dEC
= YEC µ X X, (9b)
dt
dI Ac
= YI Ac µ I Ac X, (9c)
dt
where Yγ are the yield coefficients (γ = EA, EC, I Ac) and µ I AC is the maximum isoamyl
acetate formation rate.
The common practice recommends completely removing vicinal diketones (VDKs)
since they add some undesired buttery flavor notes. VDK production is assumed to be
proportional to the growth rate, while their possible re-assimilation by yeast to form other
by-products is proportional to their concentration. It must be noticed that acetaldehyde
Foods 2022, 11, 3602 6 of 20

(AAl) is another compound showing similar behavior to the VDKs. The mass-balance
equations read:

dVDK
= YVDK µ X X − rVDK VDKX, (10a)
dt
dAAl
= YAAl (µG + µ M + µ N ) X − r AAl AAlX, (10b)
dt
where Yω defines the yield coefficients and rω defines first-order rate constants (ω = VDK, AAl).
µ X is the biomass growth rate and µG , µ M , µ N are the sugar consumption rates.
This model therefore proposes a detailed description of the flavor and aroma of
the beer but also presents the drawback of being difficult to apply in a realistic context
since it requires numerous and often expensive advanced on-line monitoring devices.
Indeed, in practice, biomass concentration is only measured at the beginning of a batch
without further monitoring. Moreover, most of the parameters are temperature-dependent,
either imposing rigorous operating conditions (i.e., only one constant temperature level)
or parameter estimation of the temperature-dependent functions (which requires data
collection at various temperatures).

3.2. Model of De Andrés-Toro et al.

This model, published in 1998, is more concise than the previous one as it only consid-
ers five state variables: sugars, biomass, ethanol, ethyl acetate, and diacetyl (i.e., vicinal
diketones) [8]. Ethyl acetate and diacetyl are assumed to be the most influencing com-
pounds regarding aroma and flavor. In the following, the model dynamics are described
state-by-state. Biomass is segregated into three types: lagged, active, and dead. It is indeed
assumed that part of the biomass goes through several states during the process, first in a
lag phase when the fermentation starts, then in an active (growing) state, and eventually in
an inactive (dead) state.

Lagged biomass
dX L
= −µ L X L , (11a)
dt
Active biomass
dX A
= µ X X A + µ L X L − µ DT X A , (11b)
dt
Dead biomass
dXD
= µ DT X A − µSD XD , (11c)
dt
where the lagged biomass becomes active at the specific rate µ L , the active biomass grows
at the specific rate µ X and dies at the rate µ DT , while the dead biomass settles in the bottom
of the reactor at the rate µSD . µ X and µSD are further defined as

µ X0 S
µX = , (12)
0.5S0 + E
µSD0 0.5S0
µSD = . (13)
0.5S0 + E

µ X represents an activation by the substrate S and an inhibition by ethanol E. The inhibition

constant is assumed to be inversely proportional to half the initial substrate concentration
S0 (indeed two units of S give one unit of E in the stoichiometry of the reaction). µSD
describes an inhibition by ethanol, which is directly related to CO2 , which is not a variable
in this model, but whose bubbles impair the settling phenomenon. The inhibition constant
is again related to half the initial substrate concentration (i.e., maximal quantity of ethanol
that can be produced).
Foods 2022, 11, 3602 7 of 20

Sugar consumption follows a Monod law according to:

dS
= −µS X A , (14)
dt
with
µ s0 S
µS = , (15)
KS + S
in which µS0 is the maximum specific consumption rate and KS is the half-saturation
constant.
Ethanol production is described by

dE
= µE X A , (16)
dt
where the specific rate includes a Monod factor with respect to the substrate S and an
inhibition factor related to the ethanol concentration. This factor vanishes when the ethanol
reaches the maximum value 0.5S0 :

µ E0 S
 
E
µE = 1− . (17)
kE + S 0.5S0

Ethyl acetate is produced as a byproduct of active biomass growth:

dEA
= YEA µ X X A , (18)
dt
where YEA is the yield coefficient.
Diacetyl is a component belonging to the vicinal diketones (VDKs), which is produced
as the biomass grows by consuming the sugars. Afterward, diacetyl is reduced into acetoin
with a reduction rate rVDK activated in the presence of ethanol:

dVDK
= kVDK SX A − rVDK VDKE. (19)
dt
All the parameters are assumed to be affected by temperature according to the Arrhe-
nius law (Equation (3)).
The biomass segregation model provides an accurate and consistent description of the
process but, as a drawback, requires the corresponding monitoring equipment. In the study
of [8], total biomass was measured online by absorbance change detection in a photocell,
while the biomass state classification was made based on pre-established assumptions.

3.3. Model of Trelea et al.

The originality of the model of Trelea et al. [9], with respect to the previous ones, is that
it considers the carbon dioxide dynamics instead of the biomass dynamics. Carbon dioxide
sensors are indeed easily implemented and calibrated on-line, reliable, and significantly
cheaper than biomass measurement devices.
The evolution of carbon dioxide is related to yeast growth, sugar consumption, and
ethanol production. CO2 dynamics are assumed to be driven by a Monod law describing
sugar activation and saturation effects, an inhibition factor taking account of the decreasing
cell respiratory capacity following ethanol accumulation. The influence of the initial
biomass concentration on the initial CO2 production rate is also taken into account:

dCO2 S 1
= µmax (CO2 + C0 X0 ), (20)
dt K S + S 1 + K I E2

where KS is the half-saturation coefficient, K I is the inhibition constant, and C0 is a conver-

sion factor.
Foods 2022, 11, 3602 8 of 20

Algebraic equations describe the evolution of the sugar and ethanol concentrations in
relation to CO2 as suggested in [12]:

S = S0 − YS CO2 , (21)

E = YE CO2 , (22)
where YS and YE are the corresponding yield coefficients. The main advantage of this
model lies in its practical and control-oriented description of the fermentation process,
considering a variable that can be easily measured, such as carbon dioxide. The maximum
specific growth rate, µmax , is however, assumed to depend on several operating parameters,
such as pressure, temperature, and initial yeast concentration, as follows:

µmax = a T Tn + a P Pn + a X X0,n + a TP Tn Pn + a TX Pn X0,n + a0 , (23)

where ai are parameters to be estimated.

3.4. Proposed Mathematical Models

In this study, two mathematical models of beer fermentation are proposed, one based
on biomass dynamics and the other on CO2 dynamics. These models take inspiration
from the mathematical developments of the previous sections and attempt to describe
experiments performed with the beer fermenter described in Section 2.
Figure 2 shows some data collected in a batch experiment at 19 °C. It is apparent
that sugar was not completely consumed at the end of the fermentation, with a residual
concentration of about 12 g/L (this behavior was confirmed in repeated experiments in
the same and different conditions). Two possible causes of this sluggish fermentation in
the end of batch were explored in additional experiments, including biomass settling and
water quality. However, gentle agitation and tests with different water sources did not
influence the initial observation. Other causes such as the depletion of some components
required for yeast proliferation and maintenance (such as nitrogen, sterols, fatty acids)
could not be assessed. The models were, therefore, adapted to describe the experimental
observations. As the published models only consider the total consumption of sugars,
structural modifications were made to cope with this type of behaviour. The changes mostly
impact the definition of the specific growth rate, as explained in the following sections.

3.4.1. Dynamic Model Based on Biomass

An overall alcoholic fermentation reaction can be written as follows:
r1
kS S −
→ k E E + kV VDK + k CO2 CO2 + X (24)

where sugars S are consumed and converted by yeasts X into ethanol E, vicinal diketones
VDK, and carbon dioxide CO2 . k S , k E , kV , and k CO2 represent the yield coefficients of sugar,
ethanol, vicinal diketones, and carbon dioxide, respectively.
Foods 2022, 11, 3602 9 of 20

100 30

Ethanol (g/l)
Sugars (g/l)
Model output
Experimental Data 20

50
10

0
0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time (h) Time (h)

Vicinal diketones
4 2
biomass (g/l)

3 1.5

(ppm)
2 1

0.5
1
0
0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time (h) Time (h)
Carbon Dioxide (l)

200

150

100

50

0
0 10 20 30 40 50 60
Time (h)

Figure 2. Direct validation of the biomass model with the data from experiment 2. Stars: experimental
data. Error bars: 95% confidence intervals. Continuous blue line: model prediction.

Vicinal diketones are later reduced by yeasts, producing 2-3-butanodiol P in the

following second reaction:
r2
VDK −→ P (25)
From Equations (24) and (25), a set of mass-balance ODEs can be derived:

dX
= µ X X − δX X, (26a)
dt
dS
= −k S µ X X, (26b)
dt
dE
= k E µX X (26c)
dt
dCO2
= −k CO2 µ X X, (26d)
dt
dVDK
= kV µ X X − rVDK VDK. (26e)
dt
The specific growth rate is defined as:
 
S
µ X = µmax 1 − min , f or S ≥ Smin , (27a)
S
= 0, f or S ≤ Smin . (27b)

The specific growth rate usually represented by a Monod law is replaced by a Droop-
like factor [13], commonly used to describe microalgae growth. This kinetic structure
expresses that a minimum level of sugar Smin is necessary to trigger growth. Above this
threshold, the Monod factor has an activation/saturation effect similar to a Monod factor.
Foods 2022, 11, 3602 10 of 20

Furthermore, a decrease in biomass is observed due to settling and is represented by the

coefficient δX in Equation (26a).
In classical batch fermentation, the temperature is the only variable that can be ma-
nipulated during the batch to control the evolution of the fermentation. Based on the
previously discussed modeling studies [7–9] and experimental observation, a temperature-
dependency of the maximum specific growth rate µmax is expected that can be described
following several structural laws [14]. Moreover, the reduction speed of the VDKs, rVDK ,
is also affected by temperature. In Table A1, the definitions and units of each parameter
are listed.

3.4.2. Dynamic Model Based on Carbon Dioxide

The yeast concentration is difficult to monitor during the batch and, most of the
time, this variable is either indirectly measured by the turbidity of the wort, or simply
by sampling and cell counting using a cytometer or a dry-weight method. Hence, there
is a strong motivation to monitor other, more accessible, process variables and develop
dynamic models describing their evolution. Carbon dioxide may be such an indicator
since it is a product of the biochemical reactions related to sugar oxidation, which provides
the necessary energy to the yeast cells to grow and leads to ethanol production if sugar is
in excess, activating the overflow metabolism [15,16]. In this study, the model of [9] was
adapted in the following way:

dCO2
= µ X CO2 , (28a)
dt
S = S0 − k S CO2 , (28b)
E = E0 + k E CO2 , (28c)
dVDK
= kV µ X CO2 − rVDK VDK, (28d)
dt
where the specific growth rate is given by:
 
S CO2
µ x = µmax 1− . (29)
KS + S Cpmax S0

The kinetic description contains a Monod law for the sugar activation and saturation
effects, and a logistic factor to mimic the observed sigmoidal evolution of carbon dioxide.
The data sets also reveal that the maximum production of carbon dioxide is correlated with
the initial sugar concentration S0 , which therefore defines the maximum CO2 level (i.e.,
the carrying capacity of the logistic model). The initial biomass concentration, which is
assumed to be known (measured) and directly correlated to the CO2 dynamics in [9], was
not used in the current study since the initial condition of CO2 was available. The rates
µmax and rVDK are assumed to depend on temperature and Table A2 lists the definitions
and units of some parameters.

4. Structural Identifiability and Observability of the Models

Identifiability globally refers to the possibility of identifying the model parameters
from the available data. A model is structurally identifiable if all the parameters can be
uniquely determined from ideal measurements of its outputs, i.e., collected in continuous
time without errors or noise, and the knowledge of the dynamic equations [17]. If this
property is not met, any further effort to estimate the non-identifiable parameters will be
vain. However, the identifiability analysis is often omitted due to the assumed complexity
of the mathematical developments required to achieve the analysis. Recently, several
methodologies and toolboxes have been developed to significantly ease the task, as reviewed
in [18]. Some of these software tools are DAISY [19], GENSSI [20], STRIKE-GOLDD [10],
and SIAN [21].
Foods 2022, 11, 3602 11 of 20

On the other hand, practical identifiability deals with the possibility of assessing
all or some of the model parameters under realistic conditions, e.g., sampled data and
measurement noise. The Fisher Information Matrix (FIM) is useful to assess practical
identifiability through a rank test condition. An ill-conditioned FIM can indicate poor
practical parameter identifiability even if structural identifiability is met.
In this work, STRIKE-GOLDD (STRuctural Identifiability taken as Extended-Generalized
Observability with Lie Derivatives and Decomposition) was used to investigate the struc-
tural identifiability of the proposed beer fermentation models. This software tool has been
developed in MATLAB® and addresses identifiability based on the concept of observ-
ability. To this end, the model is extended by considering its model parameters as state
variables with zero dynamics. The results obtained for both models indicate that structural
identifiability is ensured only when all the state variables are measured.
Another property of interest is observability, which is a prerequisite to the design of
a state observer to reconstruct nonmeasured state variables. The results of the analysis
are provided in Tables 3 and 4 for the two dynamic models. For the model based on
biomass, the analysis reveals that the set of three measurements [CO2 , E, VDK ] is necessary
to guarantee observability. The set [CO2 , E, S] shows partial observability as VDK cannot
be reconstructed but biomass X could. The carbon dioxide model requires the measurement
of VDK together with another variable (CO2 or E or S) to fulfill the observability condition.
An observer could, therefore, be designed to estimate the sugar concentration online, which
is the most expensive measurement using an online or at-line hardware probe.

Table 3. Observability analysis of the biomass model for several measurement configurations.

Measured Outputs Observable

[CO2 , S] No
[CO2 , E] No
[S, E] No
[CO2 , S, X] No
[CO2 , E, VDK] Yes
[CO2 , E, S] VDK: No X: Yes
[CO2 , S, X, VDK] No

Table 4. Observability analysis of the carbon dioxide model for several measurement configurations.

Measured Outputs Observable

[CO2 , S] No
[CO2 , E] No
[S, E] No
[S, VDK] Yes
[E, VDK] Yes
[CO2 , VDK] Yes

5. Parameter Identification Problem

Parameter identification is achieved using classical nonlinear parameter estimation
techniques [22]. The procedure considers a Weighted Least-Square (WLS) criterion, i.e.,
a weighted sum of squared differences between model predictions and experimental
measurements:
M  
J( θ ) = ∑ y(ti ) − ymodel (ti , θ )) T W−1 (y(ti ) − ymodel (ti , θ )) (30)
i =1

where J is the value of the cost function, y(ti ) is the vector of N measured variables at the
measurement instant ti (i = 1, . . . , M), ymodel (ti , θ ) is the model prediction that depends
on the set of P parameters θ to be identified, and W is a normalization matrix where the
Foods 2022, 11, 3602 12 of 20

diagonal elements are chosen as the squares of the maximum measurements values of each
component concentration. This choice allows normalization of the prediction errors, and
is particularly well-suited to a relative error model where it is assumed that the error is
proportional to the maximum values of the observed variables:

max(y21 )
 
0 ... 0
ti
max(y2 )2
 
 0 ... 0 
 ti 
W= .. .. ..
 (31)
 .. 

 . . . . 

0 0 ... max(y N )2
ti

The estimated parameter set is obtained by minimizing a cost function J(θ ) as follows:

θ̂ = arg min J(θ ) (32)

θ

To achieve the minimization of (32), a two-step procedure is implemented: (a) a multi-

start strategy defines random sets of initial parameter values to cover as much as possible
of the search field. The minimization of (32) is first performed using the Matlab® optimizer
fminsearch (Nelder-Mead algorithm); (b) the Matlab lsqnonlin optimizer is subsequently
used from the identified global minimum (i.e., the smallest local minimum identified in
the search space by fminsearch) to refine the minimization and compute the Jacobian
matrix containing the model parameter sensitivities, denoted yθ . These sensitivities can be
exploited to compute the Fisher Information Matrix (FIM) defined as:

M
FIM = ∑ yθT (ti )Ω̂−1 yθ (ti ) (33)
i =1

where Ω̂ = ê2 W is the a posteriori covariance matrix of the measurement errors, which can
be evaluated using the weighting matrix W (Equation (31)) and an a posteriori estimator of
the relative measurement error:
J∗
ê2 = (34)
MN − P
where J ∗ is the value of the cost function at the optimum, MN represents the total number
of data, and P is the number of estimated parameters θ. An estimate of the parameter
estimation error covariance matrix can then be inferred from the Cramer–Rao bound
as follows:
Σ̂ = FIM−1 (35)
From the diagonal of the covariance matrix Σ̂, the standard deviations for each param-
eter can be extracted and the corresponding coefficients of variations can be calculated as:
σ
CV = (36)
θ̂i

To achieve the estimation of the parameters of the beer fermentation models, a total
of 4 batch experiments are considered as shown in Table 2. Out of these 4 experiments,
3 are used for parameter estimation and model direct validation (experiments 2 to 4),
while experiment 1 is used for cross-validation. An important point of the current work
is the use of all the samples of experiments 2, 3, 4 to achieve the identification, including
two temperature-varying parameters, i.e., the specific growth rate µmax and the VDK
reduction rate rVDK . Previous studies have indeed demonstrated that the other parameters
do not change significantly with temperature. In addition to the stoichiometric and kinetic
parameters, the initial conditions are also considered unknown (and are therefore estimated)
since possibly corrupted by measurement noise.
Foods 2022, 11, 3602 13 of 20

5.1. Biomass Model

This model counts 8 parameters (k S , k E , Smin , gx , kV , k CO2 , µmax , r a ) to be estimated.
The dependence on temperature of the parameters µmax and rVDK , is formulated as follows:

µmax = aln( T ) + b; rVDK = cln( T ) + d; (37)

introducing the additional parameters a, b, c, d. Several nonlinear model structures have

been considered to correlate the parameters with the temperature. It turned out that the
selected logarithmic structure provides the best results.
In practice, the identification proceeds in three steps: (a) a first parameter estimation
without explicit temperature dependence (i.e., µmax and rVDK are considered constant),
(b) an estimation of the four parameters linked to the temperature dependence (the others
being fixed at their previously estimated values), and (c) a global identification of all the
parameters starting from the previous estimates.
Good practice recommends partitioning the data using a ratio of approximately 75/25
for parameter estimation (and subsequent direct validation), and cross-validation, respec-
tively. Accordingly, three experimental data sets are used in direct validation and the
remaining one in cross-validation. Since the parameter estimation procedure aims at cap-
turing information on the process in a wide range of operations, it is legitimate to include
experiments with different initial sugar and biomass concentrations and temperature levels.
Particularly, it is important to collect informative data regarding the evolution of µmax and
rVDK with respect to temperature in (37). Among the several possible data partitions, one
possible combination appears to be: experiments 2, 3, 4 for parameter estimation (and
direct validation) and experiment 1 for cross-validation. Indeed, experiments 1 and 2 are
carried out at the same temperature, but experiment 2 also includes different sugar and
biomass initial conditions. Table 5 reports the values of the estimated parameters and their
coefficients of variations.

Table 5. Parameter estimate values and coefficients of variation (CV) for the biomass model.

Parameter Units Value CV (%)

kS gS/gX 15.3 7
kE gE/gX 6.31 6
Smin g/L 13.1 11
gx h−1 1.67 × 10−2 7
kV gVDK/gX 6.51 × 10−1 10
k CO2 L 37.1 5
a h−1 4.00 × 10−1 3
b h−1 −1.1 5
c h−1 2.50 × 10−2 10
d h−1 −5.60 × 10−2 9

Figures 2 and 3 show some direct validation results, i.e., the fitting of the model to the
experimental data collected in experiments 2 and 4 together with the a posteriori error bars
on the experimental data. The model reproduces quite well the dynamics of the several
variables, even if the biomass predictions sometimes deviate from the confidence intervals
of the data, and some deviations in the VDK production are also observed in the early
hours. The coefficients of variations confirm the good estimation results, as the maximum
relative CV is 11% for the minimum substrate quota Smin .
In order to assess the model predictive capacity, cross-validation is achieved using
the dataset from experiment 1. In this case, only the initial conditions are estimated while
the parameters are kept fixed. As shown in Figure 4, the model predicts satisfactorily the
experimental data. The biomass data again has some uncertainty, which can probably be
linked to several factors such as cell counting errors, biomass mixing (to counteract biomass
settling and collect representative samples) and nitrogen limitation [23].
Foods 2022, 11, 3602 14 of 20

100 30

Ethanol (g/l)
Sugars (g/l)
Model output
20
Experimental Data
50 10

0 -10
0 20 40 60 80 0 20 40 60 80
Time (h) Time (h)

Vicinal diketones
4 3
biomass (g/l)

3 2

(ppm)
2
1
1

0
0
0 20 40 60 80 0 20 40 60 80
Time (hr) Time (h)
Carbon Dioxide (l)

150

100

50

0
0 20 40 60 80
Time (h)

Figure 3. Direct validation of the biomass model with the data from experiment 4. Stars: experimental
data. Error bars: 95% confidence intervals. Continuous blue line: model prediction.

100 40
Sugars (g/l)

Ethanol (g/l)

30

50 20

10

0 0
0 20 40 60 80 0 20 40 60 80
Time (h) Time (h)
Carbon dioxide (l)

Vicinal diketones

3
200
(ppm)

150 2

100
1
50

0 0
0 20 40 60 80 0 20 40 60 80
Time (h) Time (h)
6
Biomass (g/l)

0
0 20 40 60 80
Time (h)

Figure 4. Cross-validation of the biomass model with the data from experiment 1. Stars: experimental
data. Error bars: 95% confidence intervals. Continuous blue line: model prediction.
Foods 2022, 11, 3602 15 of 20

Vicinal diketone dynamics present a varying latency phase followed by produc-

tion/consumption, both driven by biomass dynamics. The uncertainty on the latency
period compromises the resulting fitting since biomass does not exhibit the same behavior
in the early phase of fermentation.

5.2. Carbon Dioxide Model

A similar procedure was applied to estimate the values of the carbon dioxide model
parameters. In this case, the dependence on temperature of µmax and rV DK is best repre-
sented by:
µmax = aln( T ) + b; rVDK = cT 2 + dT + e; (38)
Hence, the resulting model counts 10 parameters (k S , k E , KS , Cpmax , kV , a, b, c, d, e),
and Table 6 reports the estimated values with their respective coefficients of variations. As
can be noticed, the latter are smaller than the ones of the previous model, mainly due to
the absence of biomass measurement and the associated uncertainty.
The identification is again decomposed into distinctive steps: (a) estimation of the
parameters without temperature dependence and with an arbitrary value for KS whose
practical identifiability is poor, (b) estimation of KS with all the other parameters fixed at
their previously estimated values, (c) estimation of the parameter linked to the temperature
dependence (a to e) with all the others fixed to their previous values, and (d) final re-
estimation of all the parameters.
Figures 5 and 6 show the direct validation with experiments 2 and 4, as well as the a
posteriori error bars on the experimental data.

100 30
Model output
Experimental Data 25
80
Ethanol (g/l)
Sugars (g/l)

20
60
15
40
10

20
5

0 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time (h) Time (h)

200
2
Carbon Dioxide (l)

Vicinal diketones

150
1.5
(ppm)

100
1

50 0.5

0 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time (h) Time (h)

Figure 5. Direct validation of the carbon dioxide model with the data from experiment 2. Stars:
experimental data. Error bars: 95% confidence intervals. Continuous blue line: model prediction.
Foods 2022, 11, 3602 16 of 20

100 35
Model output
30
80 Experimental Data

Ethanol (g/l)
25

Sugar (g/l)
60 20

40 15

10
20
5

0 0
0 20 40 60 80 0 20 40 60 80
Time (h) Time (h)

200 3
Carbon Dioxide (l)

2.5

Vicinal diketones
150
2

(ppm)
100 1.5

1
50
0.5

0 0
0 20 40 60 80 0 20 40 60 80
Time (h) Time (h)

Figure 6. Direct validation of the carbon dioxide model with the data from experiment 4. Stars:
experimental data. Error bars: 95% confidence intervals. Continuous blue line: model prediction.

Table 6. Parameter estimate values and coefficients of variations (CV) for the carbon dioxide model.

Parameter Units Value CV (%)

kS gS/lCO2 3.72 × 10−1 4
kE gE/lCO2 1.62 × 10−1 3
KS g/L, 12.0 9
Cpmax L/g 2.18 3
kV ppm 1.74 × 10−2 4
a h−1 1.41 6
b h−1 − 3.89 8
c °C 2 h−1
− − 3 × 10−4 12
d °C−1 h−1 1.6 × 10−2 6
e h−1 − 1.75 × 10−1 5

Model cross-validation using experiment 1 is shown in Figure 7 and confirms the

satisfactory predictive capacity of the model, except for some minor deviations in the
evolution of the VDKs due to the presence of a time-varying latency phase.
In Table 7, the root means square errors (RMSEs) are provided for each experiment
and each variable separately. The cost function residuals of the direct validations are also
provided. It can be observed that, overall, the RMSE values are small for both models.
Regarding the biomass model, X and VDK present larger RMSEs than the other variables,
due to the observed deviations between the model prediction and the experimental data in
Figures 2 and 3. Regarding the carbon dioxide model, RMSEs indicate a better fit with the
experimental data. This statement is confirmed by Figures 5 and 6. Cross-validation results
also support this analysis since the RMSEs of the biomass and carbon dioxide models are,
respectively, 1.997 and 0.846.
Foods 2022, 11, 3602 17 of 20

Table 7. Cost function residuals and relative RMSEs for each state variable of the two models,
resulting from direct validation.

Cost Function
Model Variable Exp 2 RMSE Exp 3 RMSE Exp 4 RMSE Global RMSE
Residual
CO2 4.86 × 10−2 5.40 × 10−2 6.40 × 10−2 5.32 × 10−2
S 3.37 × 10−2 8.12 × 10−2 5.43 × 10−2 5.80 × 10−2
Biomass 1.76 E 4.78 × 10−2 6.67 × 10−2 5.15 × 10−2 5.30 × 10−2
X 9.96 × 10−2 9.63 × 10−2 1.62 × 10−1 1.09 × 10−1
VDK 1.12 × 10−1 1.16 × 10−1 1.51 × 10−1 1.12 × 10−1
CO2 3.45 × 10−2 2.99 × 10−2 4.74 × 10−2 3.56 × 10−2
S 2.23 × 10−2 7.40 × 10−2 4.63 × 10−2 5.36 × 10−2
Carbon dioxide 0.72 E 5.47 × 10−2 9.91 × 10−2 4.11 × 10−2 6.23 × 10−2
VDK 6.26 × 10−2 8.59 × 10−2 1.12 × 10−1 7.50 × 10−2

Discriminating among the proposed models is difficult since they target different
variables. However, taking into account the cost function residuals, the carbon dioxide
model fit better to the current operating conditions and monitoring set-up (J = 0.72) than
the biomass model (J = 1.76). Furthermore, from a practical point of view, the identification
of the dioxide carbon model requires a sensor configuration that is easier to set up, limiting
offline analytical analysis. Conversely, the identification of the biomass model requires
offline cell counting to measure yeast concentration. Moreover, considering process con-
trol, carbon dioxide online sensors are affordable, whereas biomass sensors are expensive
(alternatively a biomass software sensor could be developed based on the measurements
of CO2 , E, VDK). The main advantage of the biomass model lies in the provided informa-
tion about the biomass metabolic state during the fermentation process, allowing a more
straightforward detection of possible contamination.

120 40

100
30
Ethanol (g/l)
Sugars (g/l)

80

60 20

40
10
20

0 0
0 20 40 60 80 0 20 40 60 80
Time (h) Time (h)

200 2
Carbon dioxide (l)

Vicinal diketones

150 1.5
(ppm)

100 1

50 0.5

0 0
0 20 40 60 80 0 20 40 60 80
Time (h) Time (h)

Figure 7. Cross-validation of the carbon dioxide model with the data from experiment 1. Stars:
experimental data. Error bars: 95% confidence intervals. Continuous blue line: model prediction.
Foods 2022, 11, 3602 18 of 20

6. Conclusions
The demand for processes with more rigorous quality standards, as is the case in
the pharmaceutical industry, has led to the development of approaches such as process
analytical technologies (PAT), now being extended to the agro-food sector and, more
specifically, to the brewing industry. This work is motivated by the growing importance
of mathematical modeling, in the context of PAT, to design process digital twins that can
support lab-scale operations. Model-based advanced monitoring and control techniques
can indeed be developed in view of optimizing and improving the process. In this study,
two alternative models, initially proposed in seminal works, are adapted and identified
under realistic experimental conditions. One of the models is based on the description of
the biomass evolution, while the other, more pragmatic, considers carbon dioxide, a more
accessible variable that can be measured with cheap sensors. These models take account
of the temperature influence in a simple way. A systematic identification procedure is
described. Cross-validation highlights the good predictive capability of both models, which
are good candidates for model-based control.

Author Contributions: Conceptualization, J.M.Z.L., L.D., and A.V.W.; methodology, J.M.Z.L., L.D.,
H.H.E., and A.V.W.; software, J.M.Z.L.; validation, J.M.Z.L.; formal analysis, J.M.Z.L.; investigation,
J.M.Z.L., L.D., H.H.E., and A.V.W.; data curation, J.M.Z.L.; writing—original draft preparation,
J.M.Z.L., L.D., and A.V.W.; writing—review and editing, L.D., H.H.E., and A.V.W.; supervision, A.V.W.
and H.H.E.; project administration, H.H.E. and A.V.W.; funding acquisition, H.H.E. and A.V.W. All
authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Data Availability Statement: The data presented in this study are available on request from the
corresponding author.
Acknowledgments: The authors are grateful to Vincent Moeyaert, Perla Melendez, and Ivette
Navarro for their help in the experimental set-up. J.M.Z.L. acknowledges the support of CONACYT
under grant 782850.
Conflicts of Interest: The authors declare no conflict of interest.

Appendix A

Table A1. Parameter nomenclature list for biomass model.

Parameter Definition Units

kS Sugar yield coefficient gS/gX
kE Ethanol yield coefficient gE/gX
Smin Minimum sugar quota g/L
δX Settling constant h−1
kV VDK yield coefficient gVDK/gX
k CO2 CO2 yield coefficient L
Temperature-dependency
a h−1
coefficient
Temperature-dependency
b h−1
coefficient
Temperature-dependency
c h−1
coefficient
Temperature-dependency
d h−1
coefficient
Foods 2022, 11, 3602 19 of 20

Table A2. Parameter nomenclature list for carbon dioxide model.

Parameter Definition Units

kS Sugar yield coefficient gS/lCO2
kE Ethanol yield coefficient gE/lCO2
KS Substrate limitation coefficient g/L
Cpmax Carrying capacity L/g
kV VDK yield coefficient ppm
k CO2 CO2 yield coefficient L
Temperature-dependency
a h−1
coefficient
Temperature-dependency
b h−1
coefficient
Temperature-dependency
c °C−2 h−1
coefficient
Temperature-dependency
d °C−1 h−1
coefficient
Temperature-dependency
e h−1
coefficient

References
1. Iserentant, D. Beers: Recent Technological Innovations in Brewing. In Fermented Beverage Production; Lea, A.G.H., Piggott, J.R.,
Eds.; Springer: Boston, MA, USA, 2003; pp. 41–58. [CrossRef]
2. Barth, S.J. BarthHass Report Hops 2020/2021; Technical Report; BarthHaas: Nuremberg, Germany, 2021.
3. Willaert, R. The Beer Brewing Process: Wort Production and Beer Fermentation. In Handbook of Food Products Manufacturing; John
Wiley & Sons, Ltd.: Hoboken, NJ, USA, 2007; Chapter 20, pp. 443–506. [CrossRef]
4. Guido, L.; Rodrigues, P.; Rodrigues, J.; Gonçalves, C.; Barros, A. The impact of the physiological condition of the pitching yeast
on beer flavour stability: An industrial approach. Food Chem. 2004, 87, 187–193. [CrossRef]
5. Rodman, A.D.; Gerogiorgis, D.I. Multi-objective process optimisation of beer fermentation via dynamic simulation. Food Bioprod.
Process. 2016, 100, 255–274. [CrossRef]
6. Shopska, V.; Denkova, R.; Lyubenova, V.; Kostov, G. 13-Kinetic Characteristics of Alcohol Fermentation in Brewing: State of Art
and Control of the Fermentation Process. In Fermented Beverages; Grumezescu, A.M., Holban, A.M., Eds.; Woodhead Publishing:
Sawston, UK, 2019; pp. 529–575. [CrossRef]
7. Gee, D.A.; Ramirez, W.F. A flavour model for beer fermentation. J. Inst. Brew. 1994, 100, 321–329. [CrossRef]
8. De Andrés-Toro, B.; Girón-Sierra, J.; López-Orozco, J.; Fernández-Conde, C.; Peinado, J.; García Ochoa, F. A kinetic model for
beer production under industrial operational conditions. Math. Comput. Simul. 1998, 48, 65–74. [CrossRef]
9. Trelea, I.C.; Titica, M.; Landaud, S.; Latrille, E.; Corrieu, G.; Cheruy, A. Predictive modelling of brewing fermentation: From
knowledge-based to black-box models. Math. Comput. Simul. 2001, 56, 405–424. [CrossRef]
10. Villaverde, A.F.; Barreiro, A.; Papachristodoulou, A. Structural Identifiability of Dynamic Systems Biology Models. PLos Comput.
Biol. 2016, 12, 1–22. [CrossRef] [PubMed]
11. Xie, J.; Tian, X.F.; He, S.G.; Wei, Y.L.; Peng, B.; Wu, Z.Q. Evaluating the Intoxicating Degree of Liquor Products with Combinations
of Fusel Alcohols, Acids, and Esters. Molecules 2018, 23, 1239. [CrossRef] [PubMed]
12. Pandiella, S.; García, L.; Díaz, M.; Daoud, I. Monitoring the carbon dioxide during beer fermentation. MBAA Technol. 1995,
32, 126–131.
13. Droop, M. Vitamin B12 and marine ecology. IV. The kinetics of uptake, growth and inhibition in Monochrysis lutheri. J. Mar. Biol.
Assoc. UK 1968, 48, 689–733. [CrossRef]
14. Zwietering, M.H.; Jongenburger, I.; Rombouts, F.M.; van ’t Riet, K. Modeling of the Bacterial Growth Curve. Appl. Environ.
Microbiol. 1990, 56, 1875–1881. [CrossRef]
15. Crabtree, H.G. Observations on the carbohydrate metabolism of tumours. Biochem. J. 1929, 23, 536–545. [CrossRef]
16. Sonnleitner, B.; Käppeli, O. Growth of Saccharomyces cerevisiae is controlled by its limited respiratory capacity: Formulation and
verification of a hypothesis. Biotechnol. Bioeng. 1986, 28, 927–937. [CrossRef] [PubMed]
17. Bellman, R.; Åström, K. On structural identifiability. Math. Biosci. 1970, 7, 329–339. [CrossRef]
18. Wieland, F.G.; Hauber, A.L.; Rosenblatt, M.; Tönsing, C.; Timmer, J. On structural and practical identifiability. Curr. Opin. Syst.
Biol. 2021, 25, 60–69. [CrossRef]
19. Bellu, G.; Saccomani, M.P.; Audoly, S.; D’Angiò, L. DAISY: A new software tool to test global identifiability of biological and
physiological systems. Comput. Methods Programs Biomed. 2007, 88, 52–61. [CrossRef] [PubMed]
20. Chis, O.; Banga, J.R.; Balsa-Canto, E. GenSSI: A software toolbox for structural identifiability analysis of biological models.
Bioinformatics 2011, 27, 2610–2611. [CrossRef] [PubMed]
Foods 2022, 11, 3602 20 of 20

21. Hong, H.; Ovchinnikov, A.; Pogudin, G.; Yap, C. SIAN: Software for structural identifiability analysis of ODE models. Bioinfor-
matics 2019, 35, 2873–2874. [CrossRef] [PubMed]
22. Walter, E.; Pronzato, L. Idnetification of Parametric Models from Experimental Data, 1st ed.; Springer: London, UK, 1997.
23. Schulze, U.; Lidén, G.; Nielsen, J.; Villadsen, J. Physiological effects of nitrogen starvation in an anaerobic batch culture of
Saccharomyces cerevisiae. Microbiology (Reading) 1996, 142, 2299–2310. [CrossRef] [PubMed]