1. Introduction
Microorganisms are industrially used to produce either generic biomass
or specific substances such as enzymes, amino acids or antibiotics. When
biomass production is performed, usually a high yield is targeted. ForSaccharomyces cerevisiae (S. cerevisiae ) this is achieved
by purely oxidative metabolism (Madigan, Bender et al. 2017). S.
cerevisiae can use glucose as energy and carbon source for aerobic
respiration. When a critical glucose concentration is exceeded, this
microorganism produces ethanol. This is known as Crabtree effect and
results in a lower biomass yield which might be, depending on the goal
of the cultivation, undesirable (De Deken 1966). Therefore, in order to
achieve high productivity of biomass continues real-time monitoring of
ethanol concentration is required.
Bioprocess variables are of a chemical, physical, or biological nature
and can be measured in the gas, liquid, and on solid phases of a
bioprocess. On-line measurements of these variables make great demands
on the sensing device. It is easier to meet these demands in the gas
phase environment than in the liquid phase for various reasons: in the
gas phase, the number of interfering substances is smaller and the
mechanical stress on the sensor membrane is lower than in the stirred
bioreactor liquid. In addition, prevention of so-called sensor fouling
by cell adhesion is not necessary and a sterile barrier in the form of a
mass filter can be easily introduced into the gas stream (Wild, Citterio
et al. 1996).
Numerous methods attempt to measure the concentration of volatile
organic compounds (VOCs) from the vapor phase. In recent years,
particularly due to recent technological developments in sensor
technology and computing power, gas sensor arrays (electronic nose
techniques) have become valuable tools for VOC measurements. Generally,
the sensor array technique is attractive for a number of significant
features, such as the relatively fast assessment of headspace, a
quantitative representation or qualitative identification of a gas and
cheap chemical sensors which can be easily integrated in current
production processes, thus becoming particularly suitable for the
continuous monitoring of microbial fermentation processes (Jiang, Zhang
et al. 2015). Recent applications of gas sensor arrays for monitoring
fermentation process are reported in literature (Buratti and Benedetti
2016; Hidayat, Nuringtyas et al. 2018; Tan, Xie et al. 2018; Ghosh, Tudu
et al. 2017; Li, Yuan et al. 2019; Tan, Balasubramanian et al. 2019).
However, only a few works in literature demonstrate the application of
gas sensor arrays for monitoring ethanol concentration during S.cerevisiae cultivation (Mandenius, Eklöv et al. 1997; Lidén,
Mandenius et al. 1998; Bachinger and Mandenius 2001). In order to
predict a specific volatile compound with a gas sensor array,
chemometric modeling techniques are required. In the previous studies,
the calibration methods for the chemometric models are limited to
data-driven calibration methods. The main disadvantage of data-driven
calibration methods is the huge amount of off-line data necessary to
calculate a reliable model.
An alternative to data driven calibration method, which is a time
consuming task, is the model-based calibration method. A statistical
model-based approach for developing calibration models does not require
the time expensive collection of samples for off-line measurements.
Furthermore, this approach addresses some of the shortcomings of
traditional calibration methods to study the entire system response
which results in robust calibration. Lin et al. (Lin and Recke 2007)
give a systematic approach for development of data-driven soft sensors.
Model-based calibration approaches have been implemented on
spectroscopy-based monitoring systems. Solle et. al (Solle, Geissler et
al. 2003), as well as Paquet-Durand et al. (Paquet‐Durand, Assawarajuwan
et al. 2017) had used this evaluation technique for the prediction of
biomass, glucose, and ethanol during a S. cerevisiae cultivation.
Furthermore, Paquet-Durand et al. (Paquet‐Durand, Ladner et al. 2017 a)
applied this method for evaluation of fluorescence measurements during
several parallel cultivations of H. polymorpha in a microtiter
plate.
Based on fluorescence measurements Ödman et. al (Ödman, Johansen et al.
2009) and Solle et. al (Solle, Geissler et al. 2003) have evaluated
yeast cultivations using glucose as substrate and developed chemometric
models, one for the glucose consumption phase with concomitant ethanol
production and one separate for the ethanol consumption phase (after
glucose depletion). They have stated that it was difficult to use one
and the same model for both phases. Paquet-Durand et al. (Paquet‐Durand,
Assawarajuwan et al. 2017) examined artificial neural networks for the
correlation between the fluorescence spectra with glucose, biomass and
ethanol concentrations. They implemented a model-based training approach
for training the neural network with only using a single model. They
have reported an accurate prediction of glucose and biomass (error of
prediction below 5%) however the prediction error of ethanol was 10 %.
This is due to ethanol not being fluorescent and it could only be
determined indirectly from the spectra. Therefore fluorescence-based
monitoring methods are not the most accurate methods for predicting
ethanol concentrations during S. cerevisiae cultivation process.
In this contribution, ethanol concentration during yeast cultivation was
predicted using a gas sensor array and chemometric modeling. The main
contribution of this paper can be summarized as follow:
- Design and implementation of a gas sensor array and the headspace
sampling system in order to achieve accurate prediction of ethanol
concentration in the liquid phase during S. cerevisiae batch
cultivation.
- Implementation of a model-based calibration algorithm for the
calibration of the gas sensor array. Instead of using off-line
measurements, simulated process variables were used to determine
parameters of the chemometric model. The kinetic parameters of the
process model are unknown in the beginning and are also determined
during this procedure.
The results of the proposed calibration method are compared with a
classical calibration method which the parameters of the model are
acquired by least squares fitting to off-line measurements.
The remaining paper is organized as follows. Section 2 provides the
materials and methods which were applied in this study. Section 3
provides the results and Section 4 concludes this paper.