Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964

https://doi.org/10.1007/s10295-020-02308-1

BIOTECHNOLOGY METHODS - ORIGINAL PAPER

Towards smart biomanufacturing: a perspective on recent

developments in industrial measurement and monitoring
technologies for bio‑based production processes
Carina L. Gargalo1 · Isuru Udugama1 · Katrin Pontius1 · Pau C. Lopez1 · Rasmus F. Nielsen1 · Aliyeh Hasanzadeh1 ·
Seyed Soheil Mansouri1 · Christoph Bayer2 · Helena Junicke1 · Krist V. Gernaey1

Received: 22 June 2020 / Accepted: 31 August 2020 / Published online: 7 September 2020
© The Author(s) 2020

Abstract
The biomanufacturing industry has now the opportunity to upgrade its production processes to be in harmony with the latest
industrial revolution. Technology creates capabilities that enable smart manufacturing while still complying with unfolding
regulations. However, many biomanufacturing companies, especially in the biopharma sector, still have a long way to go
to fully benefit from smart manufacturing as they first need to transition their current operations to an information-driven
future. One of the most significant obstacles towards the implementation of smart biomanufacturing is the collection of large
sets of relevant data. Therefore, in this work, we both summarize the advances that have been made to date with regards to
the monitoring and control of bioprocesses, and highlight some of the key technologies that have the potential to contribute
to gathering big data. Empowering the current biomanufacturing industry to transition to Industry 4.0 operations allows for
improved productivity through information-driven automation, not only by developing infrastructure, but also by introducing
more advanced monitoring and control strategies.

Keywords Bioprocesses · Industry 4.0 · Big data · Control · Sensors · Smart biomanufacturing

Abbreviations NIR Near infrared

AI Artificial intelligence OD Optical density
BCA Background corrected absorption PAT Process analytical technologies
NAD(P)H Nicotinamide adenine dinucleotide phosphate PCA Principal components analysis
FDA Food and drug administration PCC Pearson correlation coefficient
FTIR Fourier-transform infrared spectroscopy PLS Partial least squares
GC–MS Gas chromatography–Mass spectrometry QbD Quality by design
GFP Green fluorescent protein TA Total absorption
GMP Good manufacturing practice TRL Technology readiness level
HMI Human machine interfaces UV Ultraviolet
HPLC High-performance liquid chromatography YFP Yellow fluorescent protein
MIR Mid-infrared spectroscopy YPD Yeast extract peptone dextrose
ML Machine learning
MVDA Multivariate data analysis
Introduction

* Krist V. Gernaey Industry 4.0 and the smart manufacturing movement now
[email protected] provide biomanufacturing the opportunity to upgrade
1 its production processes to be in harmony with the latest
Process and Systems Engineering Center (PROSYS),
Department of Chemical and Biochemical Engineering, industrial revolution [118]. Due to the promise of increased
Technical University of Denmark, Kgs. Lyngby, Denmark productivity and flexibility, there is significant interest from
2
Department of Process Engineering, TH Nuremberg, both managers and process engineers to transform their
Nuremberg , Germany plants to smart manufacturing facilities. However, in many

13
Vol.:(0123456789)
948 Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964

companies, manufacturing is usually one of the last parts to the collection of large sets of relevant data. The current situ-
embrace innovation, since doing so might come with great ation is that many companies are still in the process of shift-
investment, expensive downtime, and prolonged licensing ing from manual to automated systems (Industry 3.0) [4].
and regulatory updates. Facing this risk, management has Hence, there is still a long way to go for many biomanu-
to be convinced of the real and distinct benefits that can be facturing companies to fully benefit from smart manufac-
achieved by implementing real innovation on the produc- turing as they need to transition their current operations to
tion floor. a data-rich future first. Thus, the objective of this manu-
One of the trademarks of Industry 4.0 is big data, which script is both to summarize recent advances with regards
refers to large sets of process and product data collected to the monitoring and control of bioprocesses, and to high-
by sensors and process analytical technologies (PAT) [1]. light some of the key technologies that have the potential to
Among several other benefits, integration of data from oper- overcome the aforementioned limitations, which have, so
ations and business activities can promote productivity by far, prevented smart manufacturing from being fully real-
allowing greater visibility across upstream and downstream ized in the bio-based industry. By developing the required
operations. Being able to use historical and real-time data infrastructure as well as data-driven monitoring and control-
to predict future outcomes is an empowering tool that can strategies, this transition of the current biomanufacturing
help employees to be proactive instead of reactive. They industry to Industry 4.0 operations allows for improved pro-
can understand in an agile manner what is happening in a ductivity while ensuring regulatory compliance.
process and why, as well as predict what will happen when
variations occur [2]. The gains obtained from such a proac-
tive, predictive feed-forward control approach can exceed State of the art
the incremental yield improvements that companies seek.
Besides bringing transparency and more informed deci- Smart factories hold the promise to also increase sustain-
sions, being able to use these data will also allow the factory ability through real-time monitoring of production, where
to, for example, easily adapt to schedule and product changes the automated control systems are expected to reduce the
in a way that requires minimal intervention, allowing swift number of faulty batches and cut the maintenance costs.
changeovers and decreased cleaning validation times. Thus, the aptitude of biomanufacturing companies to auto-
Furthermore, these data can be used for optimization matically and appropriately control the bioprocesses in their
purposes by applying advanced big data analytics. Machine optimal state is of crucial importance, as this helps to reduce
Learning (ML), a branch of Artificial Intelligence (AI), is or maintain the production costs and increase yields, while
one of the ways to achieve this. ML works with small to keeping the uniformity of product quality. The technology
large datasets by analyzing and comparing the data so as to readiness level (TRL) of advanced sampling methods and
find mutual patterns and explore differences [3]. automated measurement techniques may significantly reduce
Being able to rely on automated systems that require the time needed for process monitoring and control. This is
minimal human/manual intervention will result in higher emphasized by the FDA, which highlights the use of new
yields and quality, alongside with decreased costs and waste monitoring and control approaches, such as Process Ana-
generation, which is of great importance to bio-based pro- lytical Technology (PAT) tools, to improve and guarantee
duction, and especially biopharma. product quality, particularly in the pharmaceutical industry
Although hesitant to implement AI/ML techniques due to [5]. Noteworthy is that, a great part of the PAT objectives are
strict requirements for GMP compliance [4], the biomanu- general in nature and thus can be applied not only to pharma
facturing industry has a lot to benefit from data analytics. but also to any biomanufacturing process [6].
Data analytics are the key to provide real-time insights, as Carrying out the effective implementation of PAT, or
well as enabling evaluation and validation of all critical any other monitoring and control approach, depends on,
process parameters against regulatory guidelines, ranging among other things, the availability of robust and reliable
from raw materials to the finished product. This actually sensors, and full understanding of the intrinsic variability
helps companies, especially within the biopharma sector, to of bioprocesses [7]. Thus, this framework requires process
comply with the strict and compulsory requirements that are understanding based upon scientific knowledge aiming at
characteristic of that sector. monitoring and control of all critical process parameters
Nevertheless, despite this positive outlook, there are tech- that affect the quality of the final product [8]. To this end,
nical, economic and organizational challenges, and likely PAT consists of tools that include design of experiments,
some unknowns, that must be addressed to successfully bioprocess modelling, multivariate data analysis and sensor
implement these technologies in biomanufacturing. technologies. Studies presented along the years, such as [9,
However, the most significant obstacle towards the imple- 10], have shown that there are great benefits behind devel-
mentation of smart manufacturing in bio-based industries is oping mathematical models especially for the optimization

13
Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964 949

and control of bioprocesses. The development and imple- with a considerable variability due to the inherent nature
mentation of modelling strategies, real-time monitoring, of the cell cultures. Hence, sensors are required to measure
optimization and control, is required to ensure operational physical variables such as temperature and pressure, chemi-
reproducibility, quality control and consistency [11]. How- cal quantities such as pH and dissolved oxygen as well as
ever, although being aware of the potential benefits, in the biological parameters such as cell density or metabolite con-
biomanufacturing industry the processes are still vastly centrations (Fig. 1). It may be noted that most biological
optimized and controlled without the explicit use of these variables are particularly difficult to measure and monitor.
models. Monitoring methods and their associated sensors and ana-
lyzers, can be further categorized according to their position
Monitoring and control of bioprocesses regarding the process unit as in-line, at-line or off-line as
illustrated in Fig. 2.
Notwithstanding the different objectives and end products, An in-line sensor produces data continuously (no sam-
most biomanufacturing processes include the cultivation of pling), and it is in direct contact with the process medium
microorganisms, which implies a process consisting of com- (invasive) or separated from the medium by a glass window
plex chemical, physical and biological phenomena [12]. A (bypass, non-invasive, also called on-line sensor). By pro-
dependable and consistent analytical system is necessary to viding continuous information, these sensors are enablers of
control the process conditions in all parts of the biomanu- continuous process control.
facturing process (upstream, downstream and product for- At-line sensors analyze samples in close proximity to the
mulation). For example, the upstream processing part, once bioreactor. Even though the samples are collected at regular
it includes cell growth, is a complicated multi-phase system time-intervals (manually or automatically), time delays due

On-line head space sampling

In-line
Off-gas
Air, O2, CO2
Pressure
Viscosity
Stirrer speed In-line real time (in situ)
Temperature NIR
pH DS
At-line FS
pO2 FTIR
Tens of minutes
Raman
Assay instrumentation:
PTR-MS Off-line hour(s)
GC-MS On-line sampling Sampling
Assay:
MALDI-TOF-MS
HPLC
NMR
ELISA
FIA
GC
Flow cytometry MS
Electrophoresis
Microscopy
DATA
concentrations:
biomass, glucose, lactate, Model driven
amino acids, enzymes, Soft-sensors
antibodies, organic acids,
Spectra
vitamins, recombinant Data driven Chemometrics
proteins, volatile organic
compounds, etc.

Fig. 1  Schematic of bioprocess monitoring: variables and different raphy, MS mass spectrometry, PTR-MS proton transfer reaction mass
analytical techniques. NIR near-infrared spectroscopy, DS dielectric spectrometry, MALDI-TOF-MS matrix-assisted laser desorption ioni-
spectroscopy, FTIR Fourier-transform infrared spectroscopy, FS fluo- zation time-of-flight mass spectrometry, NMR nuclear magnetic reso-
rescence spectroscopy, HPLC high-performance liquid chromatogra- nance, FIA flow-injection analysis. Not all methods can be deployed
phy, ELISA enzyme-linked immunosorbent assay, GC gas chromatog- in an on-line fashion. Adapted from [11]

13
950 Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964

using methods such as HPLC or GC/MS [19]. A few exam-

ples include in-situ biosensors [20–22], optical sensors [23,
24], at-line implementations of traditional methods (HPLC,
In-line probe Bypass GC/MS) [19, 25] and spectroscopic sensors [26–42], among
others.
In-line On-line These combined analytical and data-driven approaches
are promising new developments that have the potential
Filtration
to monitor key process parameters in real-time. There-
probe/sampling fore, they will be discussed in more detail in the following
unit sub-sections.
Analyzer
In‑situ biosensors
At-line Off-line
Biosensors are analytical tools consisting of an immobilized
sensing material in close contact with a suitable transducer
Fig. 2  In-line, at-line and off-line sensors. Adapted from [8] that converts the biochemical signals into quantifiable elec-
trical signals [43]. A typical biosensor includes three parts,
to the analysis (depending on the equipment) render such an immobilized biological detection element on a signal
data suitable for monitoring purposes, but not for control. transducer unit, which is amplified by a signal conversion
Finally, samples for off-line measurements are collected unit.
manually or automatically and then transferred to the labo- The analytes are sensitively and selectively recognized by
ratory to be analyzed. This causes long time delays, so that the bio-components either via a catalytic mechanism (e.g.
such measurements cannot contribute to the control of the enzymes, cells, tissues, organelles, etc.) or through binding
dynamic process behavior [8, 12]. (e.g. protein channels, antibodies, nucleic acids, etc.). The
interaction of the biological detection element and the ana-
Towards real‑time data collection lyte is determined by the transducer unit. This unit can use
optical [44], electrochemical [45], calorimetrical [46], or
Real-time monitoring of bioreactors is seen as a crucial piezoelectrical principles [47]. The integration of biological
part of effective bioprocess control since it can help achieve elements such as enzymes, microorganisms, and antibodies
high efficiency, productivity and reproducibility [12, 13]. as sensing materials makes the transducer selective and sen-
Although seldom available, timely process information sitive. Due to their properties, these sensors have the capa-
about biomass, substrates, metabolites, products and nutri- bility of providing fast, cost effective and reliable analytical
ents is vital to take effective control decisions [11, 14, 15]. results [43], which together with expanding computation
Furthermore, measuring products such as cells, proteins and power, enables the use of biosensors as another promising
by-products is crucial to guarantee that the production pro- solution enabling the drive towards smart manufacturing.
cess is developing as expected. The opportunity to assess Recently, several biosensors have been developed for the
how the fermentation is progressing makes it possible to not control and monitoring of many different analytes [44, 48,
only optimize the process by promptly regulating specific 49]. In particular, biosensors for the on-line determination of
parameters, but also to spot when the process is not under glucose, glutamate, and lactate are available [49, 50], which
optimal operation [11, 16]. are delivered ready to use and can be directly integrated with
In the particular case of fermentation processes, addi- bioreactors via standard ports. However, these sensors are
tional measurements may be used to measure variables such still in the development stage and are not used in industrial
as dissolved ­CO2 (in the off-gas) and optical density. For scale production processes.
example, in the case of cell density, it is usually measured
using optical in-situ probes, or chemometrics coupled with Spectroscopic sensors
intact cell mass spectrometry (ICMS) [16], impedance spec-
troscopy [17], near infrared spectroscopy (NIR) or Fourier- Spectroscopic sensors are quite promising for successful
transform-infrared spectroscopy (FTIR) [18]. bioprocess monitoring for two reasons. Firstly, nearly all
Furthermore, combined analytical and chemometric biological, chemical and physical variables are accessible by
approaches are being developed to monitor the substrate, spectroscopy using the whole spectrum (from UV to MIR)
metabolites and products in bioprocesses. Chemometrics [51]. Secondly, as non-destructive in-line sensors they pro-
are employed to provide real-time analysis of variables that vide information with little or no time delay, thus enabling
otherwise would require off-line analysis and interpretation real-time monitoring and control of several process variables

13
Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964 951

[12, 51, 52]. A good example of the implementation of spec- used to characterize correlations between process variables
troscopic sensors in biomanufacturing is the in-line use of and spectral data [8], and thus different variables can be
2-D fluorescence spectroscopy which was applied in a high- predicted on-line from these measurements. This provides
throughput fermentation system called ­BioLector®. This a thorough view of the process and enables process automa-
system enables the monitoring of different microorganisms tion and rapid fault detection by closed loop control [8].
and mammalian cells, fluorophores (e.g., GFP, YFP), and For example, in the case of fermentation, in regression
NAD(P)H [53, 54]. models such as PLS, accounting for the linear relationship
Even though spectroscopic sensors fall in line with the between the spectral data and the concentrations is very
PAT initiative [55] and seem to be a promising step towards important due to the fact that many highly correlated com-
smart manufacturing, there are far more applications of pounds contribute to the spectral matrix. Thus, by constrain-
spectroscopy published in research than in industry. There ing the I/O relation to a linear system, it is possible to make
are several reasons behind this: (i) high requirements for PLS models predict on causal relation, i.e., the absorption
well-documented analytical systems especially in GMP peaks of each pure compound are reflected in the latent
manufacturing environments; (ii) many industries do not structures of the PLS. This is important when the model is
publish the details of the monitoring procedures used; and, expected to be used in systems were these correlations can
(iii) great investment involved in its industrial implementa- change.
tion, both in terms of equipment and in terms of the require- After the model calibration, external validation is neces-
ment of highly qualified personnel to set up and maintain sary so as to ensure that the developed model is as accurate
such an instrument. Some companies have attempted to as possible and behaving as expected. The model developed
tackle the need for qualified personnel in the industries by should adapt to the process and measurement conditions,
developing solutions that give a third party access to the thus changes in the actual process such as raw materials or
data, such as AnalyticTrust from Q-Interline [56], where the equipment require model updates [8]. Model changes and
quality of the analytical instrument and the data it generates updates require highly qualified personnel, and therefore,
is monitored by a third party. However, the security issues third parties such as Analytic Trust [56] are often seen as
are a usual argued obstacle as to fully embrace these new quite relevant and necessary to provide these resources.
technologies, and it is important to ensure that all data are As highlighted, new opportunities for automation, moni-
protected sufficiently. toring and control are closely related to the use of machine
learning algorithms to interpret the large quantities of data
Chemometrics (multivariate data analysis) collected and return intelligent information.

Spectroscopic sensing produces large datasets and thus Free‐floating wireless sensors
requires chemometrics, or multivariate data analysis
(MVDA), to be able to provide continuous real-time moni- Recently, floating sensor devices without a physical connec-
toring of the bioprocess variables. MVDA is a subset of tion to a reactor have been proposed. These floating sensors
machine learning algorithms that deals with multiple vari- follow the flow in the bioreactor and collect data along a
ables simultaneously. MVDA is used to extract information trajectory, transmitting it by wireless technology [64]. Zim-
from the spectra, by processing the data and reducing the mermann et al. [65] developed the wireless spherical par-
complexity of a data set. To this end, data pre-processing is ticles for the determination of fluid dynamics. This type of
a powerful tool that can reveal relevant information in the sensors are already commercially available (e.g., smartINST,
data [8, 12, 57]. After pre-processing, model calibration is smartCAPS, France, and Freesense, Denmark) in the form
performed so as to retrieve qualitative and quantitative infor- of spheres. These spheres are deployed inside the bioreactor
mation from the spectral data. Many MVDA approaches and move around freely to measure variables such as pH,
are based on Principal Component Analysis (PCA), which temperature, pressure, conductivity, and turbidity during
is often used to investigate the structure, variance and/or production. The sensor spheres are composed of a steriliz-
distribution of the dataset and to identify outliers [58]. As able shell, electronic boards, a battery, and the sensing ele-
a qualitative approach and consistent with the PAT initia- ment. They gather data on several variables (e.g. tempera-
tive [59], PCA has been used for process supervision, to ture, pH, pressure, dissolved oxygen) simultaneously and
classify raw materials and batches, as well as to deduce the then transmit this information in real-time to an external data
process status based on the spectral data [58, 60]. In this analysis unit. These mobile wireless sensing devices offer
way, a process target line or trajectory, also called golden significant potential for real-time monitoring and control of
batch, can be identified from similar and ideal process runs bioprocesses, not only due to the on-line data collection and
[61–63]. Quantitative models, most often using partial least analysis, but also due to the fact that these data represent real
square regression (PLS) or sometimes neural networks, are gradients inside the reactor. While these sensors would not

13
952 Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964

be used during all production campaigns, they can be used biomass assessment and quantification are mostly limited
during selected ones to monitor the spatial evolution of a to off-line analysis. Even if on-line growth detection sys-
reactor (fermenter) and hence provide valuable insights so tems may be applied, their performance is impaired dur-
as to carry out operational changes. ing the very early and late stages of the fermentation, when
very low and very high biomass concentrations are present,
Soft sensors respectively.
The current status-quo method for growth detection is
Soft sensors are advanced process monitoring systems, the measurement of the optical density (OD). The increase
which use algorithms to assess measurements in an on-line in absorbance, typically measured at 600 nm, represents the
manner to generate information about an otherwise unmeas- progress of the biomass formation. However, such measure-
ured process state [6, 66]. Recently, they have been the focus ments are compromised by a narrow linear OD detection
of many studies since they appear to be an alternative to the range of 0.1–1, which equates to a cell dry weight concen-
traditional automated approaches, enabling the monitoring tration span of approximately 0.1–1 g/l. Besides, the lim-
of state variables that indeed affect the bioprocesses but can- ited reproducibility between different spectrometers also
not be measured in real-time [67–70]. Spectroscopic sensors affects the results obtained by such measurements. To this
have, in some occasions, been labeled ‘soft sensors’ in the end, there is an apparent need to identify alternative meth-
bioprocessing literature due to the fact that spectroscopic odologies that provide better detection of these critical state
data is modeled using software and these models produce parameters to make smart manufacturing a reality.
information similar to hardware sensors [57, 66, 71]. Recent progress in microscopic imaging and image
Three modelling approaches can be used for the design of analysis have opened an optical window into the reac-
soft sensors: (i) mechanistic models based upon first princi- tor, so that cellular events may be observed and microbial
ples (white box) [10, 72–74]; (ii) data-driven models (black growth be evaluated automatically. Microscopy has grown
box) [58, 61, 75, 76]; and, (iii) hybrid models (grey box) into an efficient tool establishing a basis for novel image-
[77–81]. based monitoring and future control strategies. The imag-
The growing computer capacity and the advances in sig- ing of living cells has yielded tremendous insights into
nal processing (AI and ML algorithms) have made soft sen- cellular growth, functions and responses to environmental
sors very convenient and enticing to monitor and control changes, for instance through information about cell size,
industrial manufacturing. Hence, the opportunities behind shape, position and motility [82–84]. Challenges arising
their implementation perfectly align with the PAT initiative from different microscopic techniques such as bright-field
and the smart manufacturing movement. Examples and more [85] and fluorescence microscopy [86] that result in poor
details on the development and implementation of soft sen- counting statistics, for instance a small field of view or the
sors in the biomanufacturing industry are given in [6, 66]. visualization of objects that are transparent on the image, so
called ‘phase objects’, are solved by phase contrast [87] and
confocal [88] microscopy. Yet, they are expensive, require
Enabling smart biomanufacturing dedicated infrastructure, have long acquisition times and, in
addition, improper settings can cause significant artefacts
As illustrated in the state of the art section, there are recent on the images [89].
developments in sensors, process monitoring and closed The oCelloScope instrument (BioSense Solutions ApS,
loop process control that are enabling the drive towards Farum, Denmark) is a novel and compact solution based
smart manufacturing in the biotech industry. The objective on bright-field, digital [90] and time-lapse [91] microscopy.
of this section is to present three specific examples of smart The instrument facilitates a magnification factor of 200 and
solutions developed by the team of authors at the Technical hence, enables the detection and segmentation of objects
University of Denmark in close collaboration with industrial with a size between 0.5 µm and 1 mm. Accordingly, it is suit-
partners. able for the investigation of mammalian cells, yeast/fungi,
bacteria and crystals in (semi) transparent substances. The
Case study 1: monitoring of fermentation processes software for operation and analysis includes several state-of-
by imaging and image analysis the-art algorithms facilitating automated image acquisition
and robust analysis. Objects are identified from segmented
The assessment of microbial growth and the quantification images by their true shape, while key properties such as
of biomass, respectively, is the most evident way to evaluate surface area, perimeter and circularity are assigned.
the progress of any fermentation. The microbial biomass The open software platform in use has been shown to
as such is the producing core of any fermentation process provide considerable advantages in several research fields
and frequently, the biomass is the product itself. However, including microbiology [92], medicinal chemistry [93],

13
Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964 953

pharmaceutical biotechnology [94] and basic cancer research bright spot feature was developed and integrated into the
[95]. Moreover, the technology has found a solid application software, exploiting the fact that the yeast cells appear as
area in both monitoring of bacterial growth and growth inhi- a bright spot surrounded by a dark border on the images.
bition [96], respectively, as well as for detection of changes By counting the number of bright spots associated to a
in microbial morphology [97]. cell object, the bright-spot algorithm allows the automatic
The ability to measure microbial growth and to detect distinction between single cells, budding cells and cell
morphological features simultaneously render this system clusters. An image of a yeast culture with different cell
particularly attractive. For example, information about the objects, acquired with the oCelloScope instrument, is
cell size and cell size distribution in yeast cultivations have depicted in Fig. 3.
been shown to be correlated with the cell viability (dead/ The bright spot feature was demonstrated to work reliably
alive, osmotically stressed [98, 99]) and the growth rate of with a failure rate of less than 5%, on average, as shown in
the culture [100]. Furthermore, the cell size was recently Fig. 4.
correlated to the accumulation of an internal product (fatty The validation results shown in Fig. 4 are based on a lab-
acids) in microalgae [101]. Consequently, image analysis scale yeast fed-batch fermentation experiment (2 l working
seems to be a powerful tool that provides snapshots of the volume, YPD medium, grown at 30 °C, 800 rpm, 1 vvm,
physiological state through the assessment of morphological controlled at pH 6), with glucose addition (100 ml of a
features. Ultimately, image analysis can be used to assess 400 g/l glucose solution) after 6 h. In parallel to the images
the most crucial parameters such as microbial growth, sub- collected with the oCelloScope instrument, samples from
strate and product levels at a specific point in time, which the reactor were withdrawn manually and diluted to an OD
forms the basis for novel image-based control strategies for of approximately 0.1. The yeast cell concentration corre-
fermentation process operations. lated to this OD value was previously investigated to yield
Thus far, the application has been limited to micro- an appropriate image quality, with respect to the separation
scopic slides and microtiter plates allowing off-line sam- of cell objects. On average 1300 cell objects were segmented
ple analysis only. However, the recent development of a on each image, and the results of three images [after inocula-
flow-through cell as an alternative sampling device by tion (0 min of fermentation time), 240 min and 560 min of
ParticleTech ApS enables real-time, automated on-line fermentation time] are displayed in Fig. 4. After automated
image data acquisition and analysis. Hence, by means of classification into groups of cell objects exhibiting 0, 1, 2,
the developed prototype flow-through cell, a very first trial 3, 4, and 5 (or more) bright spots, the groups were manually
of an on-line, image-based monitoring approach of a yeast screened for false objects and any false object was manually
fermentation process was set up by connecting the oCello- excluded from the respective group.
Scope via a recirculation loop to the fermenter [102]. This Furthermore, the bright spot feature is capable of exclud-
first trial aimed at both growth detection and the evalua- ing image artefacts that result from shadings or out-of-focus
tion of morphological trends. For automated distinction of cells in the background, which result in objects with zero
single cells, budding cells and cell clusters, the so-called bright spots.

Fig. 3  Image of a yeast cell

culture possessing an OD
value of approximately 0.1,
grown in YPD medium. Yeast
cells appear as a bright spot
surrounded by a darker border.
The bright spot feature counts
the number of bright spots per
object. As indicated in the fig-
ure, one bright spot represents
a single yeast cell, two bright
spots represent a budding yeast
cell and three and more than
three bright spots represent a
cluster of yeast cells

13
954 Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964

Fig. 4  Evaluation of the bright spot feature exemplarily shown on were manually selected and excluded due to false segmentation. 28%
images acquired after 0, 240 and 560 min of a lab-scale yeast fer- of all objects segmented on the image were automatically classified
mentation process. The figure at time point 0 (inoculation) is to be to have one bright spot, out of which 2% were manually selected and
interpreted as follows: 8% of all objects segmented on the image were excluded due to false segmentation. The other bars in the bar charts
automatically classified to have zero bright spots, out of which 0% have to be read accordingly

The application of the prototype flow-through cell effec- with the help of the prototype flow-through cell are shown
tively demonstrated the use of this technology for auto- in Fig. 5.
mated on-line growth detection. Images were collected However, the chosen settings (300 µm depth of the pro-
every 10 min and a selection of these images acquired totype flow-through cell) limited the upper detection to an

Fig. 5  60% zoom into the images acquired on-line, and the relative from the border of an air bubble which was temporarily stuck inside
time point of image acquisition is indicated for each image. The big the flow-through cell
shading on the left corner on the image acquired at 480 min resulted

13
Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964 955

OD value of 4 (a cell dry weight of ca. 2 g/l). The growth Case study 2: real‑time particle monitoring
detection achieved on-line by the so-called TA (total
absorption) and BCA (background corrected absorption) In the last decade, there has been significant developments
algorithms is shown in Fig. 6. Note that the growth detec- within digital imaging, image analysis algorithms and com-
tion algorithms are based on absorption measurements putational processing power. This has allowed for the devel-
(pixel intensity). Hence, reducing the liquid depth inside opment of new real-time direct particle analysis methods,
the flow-through cell decreases the overall absorption and where high-resolution microscopic imaging can be used to
thus increases the upper detection limit of biomass that capture images of particles in liquid suspension. By applying
can be measured. an image segmentation algorithm and subsequently analys-
For that reason, a new generation of flow-through cells ing the identified particles, one can obtain information on
was developed whose depth can be adjusted automati- particle population properties in a matter of seconds, includ-
cally, allowing the system to handle a wider range of cell ing particle shape-, size- and morphology-distributions. The
concentrations during the fermentation process. Secondly, process of particle analysis using image analysis can be seen
a dual pump flow controller for automatic dilution was illustrated in Fig. 7.
developed (ParticeTech ApS). In this way, samples from A number of commercial particle monitoring solutions
the fermenter can be automatically diluted providing an have become available in the last two decades. This includes
appropriate cell-concentration or, respectively, image in-line probe-based sensors (Mettler Toledo ParticleView
quality for segmentation. This, together with the height [103], SOPAT [104]) but also non-invasive on-line flow-cell
(depth) adjustment of the flow-through-cell, simplifies based sensors (Sympatec [105], ParticleTech [106]). The lat-
tremendously the acquisition of on-line image data over a ter technology has the benefit of a more controllable imaging
much larger range of cell concentrations. environment that improves image quality, which also heavily
Recent technological advances such as the oCelloScope affects the measurement accuracy. This includes decreased
instrument deliver information about both, the cell con- background noise due to thinness of the flow-cell, improved
centration and the morphology dynamics regarding cell lighting sources etc. Many of the commercial solutions have
size, the cell size distribution and the distribution between a lower detection limit down to a particle size of 0.5 μm.
single cells, budding cells and cell clusters. The latter may This allows for studying various chemical and biochemical
for example be correlated with the production of insulin, processes that have particles in suspension. For instance, it is
which is highly relevant for insulin production processes possible to study eukaryotic cells in fermentations, crystals
based on yeast. Besides, studying the effect of relevant in crystallizations, flocs in flocculation etc. With sampling
process events frequently challenging process operation, times of less than a minute, one can now observe the process
such as failure of stirring or aeration, on the cell morphol- dynamics related to the particles and furthermore use the
ogy might bring significant benefits for integrated trouble- particle analysis as data source in process control strategies.
shooting. Such findings may lead to novel, image-based Nielsen et al. [107–110] have recently analyzed a number
monitoring strategies at production scale. of different particle processes, including a lab-scale crystal-
lization, an industrial scale crystallization and a lab-scale

Fig. 6  The increase of biomass followed by a off-line OD600 measurements, and b on-line via the normalized BCA and TA algorithm during a
lab-scale yeast fermentation process (2 l working volume, YPD medium, grown at 30 °C, 800 rpm, 1 vvm, controlled at pH 6)

13
956 Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964

Area: 5265.70 µm2

EQPC: 81.88 µm
Circularity: 0.79
FeretMin: 60.29 µm
FeretMax: 124.47 µm
FeretMean: 98.52 µm

Fig. 7  The process of image analysis, from imaging (left), through segmentation (middle) to analysis of individual particles (right). The particles
are given random colors to indicate the particles detected during image segmentation

flocculation. Here they have used on-line and at-line sam- the left in Fig. 8 it is also easy to see how the crystals are
pling respectively, and analyzed the particles using the Parti- growing throughout the batch crystallizations, starting from
cleTech solution [106]. For the on-line measurements, liquid an almost uniform distribution to a wide distribution.
samples were withdrawn from the reactor using a peristaltic As the mentioned particle analysis sensors can provide
pump to an 800 µm thick flow-cell inside the Particle Tech relatively frequent measurements, it becomes possible to
microscope unit. After imaging, the liquid samples were capture intermediate process dynamics. This also makes it
returned to the reactor. For at-line measurements, samples feasible to use more data-driven and complex kinetic models
were withdrawn manually from the process tank to a micro- to improve the model prediction accuracy.
titer plate, which was then placed in the microscope unit Nielsen et al. [107, 108] have recently proposed a
for imaging. By sampling every 4–5 min, it was possible to hybrid modelling approach that accommodates the
capture the process dynamics as a set of time-series data of increased quantity of data available from real-time parti-
particle properties and particle concentration. cle analysis sensors. Here they suggest a model structure
Data from a lab-scale batch cooling crystallization, pre- combining first-principles mass and population balance
sented in the work by Nielsen et al. [103], is illustrated in models with a data-driven neural network model for esti-
Fig. 9. Two batch crystallizations were carried out with mating the process kinetics. Their model structure can be
slightly varying cooling profiles, resulting in varying particle seen in Fig. 9. The inputs to the data-driven model consist
size distributions. This can clearly be seen in the differences here of multi-dimensional data from additional at-line/on-
in median diameters (D50 FeretMean) at the end of the two line/soft-sensors. The output of the data-driven model is
batches, illustrated in the plot to the left. From the plot to a number of kinetic rates of particle phenomena, such as

30
70 Bat ch 1 t = 00:00
Relat iv e v olum e densit y [ -]

0 .5
Bat ch 2 t = 0 1 :1 9
D50 Feret Mean [ µm ]

25
Tem perat ure [ êC]

60 t = 0 2 :0 4
0 .4
t = 0 2 :3 7
50 20 t = 0 3 :1 3
0.3
t = 04:04

40 15 0 .2

0 .1
30 10

0 .0

Feret Mean [ µm ]
Tim e [ h h :m m ]

Fig. 8  Particle size attributes from a lactose cooling crystallization, Right: Relative volume density of crystals in batch 2 for selected
using data from Nielsen et al. [108]. Left: Reactor temperature and samples during the batch crystallization
median (D50) Feret Mean particle size for two batch crystallizations.

13
Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964 957

Fig. 9  Hybrid model structure

by Nielsen et al. [107] where
they have used a deep neural
network as −the machine learn-
ing model. x represents the
state variables, z represent the
−
−
control actions and y represent
the kinetic rates. Reprinted
from Computers and Chemical
Engineering, Volume 140, RF
Nielsen, N Nazemzadeh, LW
Sillesen, MP Andersson, KV
Gernaey, SS Mansouri, Hybrid
machine learning assisted mod-
elling framework for particle
processes, 106,916, 2020, with
permission from Elsevier

with only limited process knowledge and/or lack of meas-

0.30 Hybrid model prediction ured process variables.
Conventional model prediction The hybrid model structure was implemented by Nielsen
Relative number density [-]

0.25 Measured distribution et al. [108] using Tensorflow [111], an open source python
software library. Automatic differentiation was here
0.20
employed to significantly speed up the training of the hybrid
0.15 model, which opens up for training the hybrid model during
process operation, utilizing the latest process data.
0.10 Nielsen et al. [108] compared the hybrid model struc-
ture in Fig. 9, using the lactose case study data presented
0.05
in Fig. 8, where temperature was used as the only measured
0.00
process variable to model nucleation and growth rates, with
20 40 60 80
a conventional, mechanistic model. They trained/fitted the
Particle size, FeretMean [µm] hybrid model and a corresponding conventional crystalliza-
tion model respectively using data from batch 2 and carried
Fig. 10  Conventional model compared to hybrid model [108]. out end-to-batch simulations of the validation batch (batch
Reprinted from Computers and Chemical Engineering, Volume 140, 1) using both models. The resulting predictions can be seen
RF Nielsen, N Nazemzadeh, LW Sillesen, MP Andersson, KV Ger- in Fig. 10.
naey, SS Mansouri, Hybrid machine learning assisted modelling
Nielsen et al. [108] concluded that the hybrid model
framework for particle processes, 106,916, 2020, with permission
from Elsevier was capable of capturing slightly more of the phenomena
dynamics than the conventional model. There were however
still discrepancies between the hybrid model and the final
nucleation, growth, shrinkage, agglomeration and break- measured distributions, which was explained by the lack of
age rates. These rates are then included into a discretized training data. They also presented two additional case stud-
first principles population balance model. ies, on flocculation/breakage of silica particles and a phar-
The neural network here substitutes the conventional maceutical crystallization respectively, which both showed
case-specific kinetic expressions. The conventional kinetic good results where only limited prior process knowledge
expressions typically contain one to five model parameters was available.
that are estimated using small amounts of experimental One should note that the model complexity still needs to
data, and typically only rely on a few process variables. be decided based on the quality and quantity of experimental
For instance, for a crystallization such as the one presented data, as an overly complex model will lead to over-fitting and
in Fig. 10, one would typically only use the relative super- a too simple model will lead to under-fitting. The increased
saturation, calculated based on the reactor temperature and data coming from the new real-time monitoring methods do
solute concentration. Using a neural network instead, the however allow for using models with increased complexity
number of input process variables can easily be extended without over-fitting, which in the end results in higher pre-
beyond two process variables, but also reduced for systems diction accuracies.

13
958 Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964

Case study 3: monitoring and closed loop control cerevisiae detoxifies the inhibitors present in lignocellulosic
of cellulosic fermentations hydrolysates. All these changes in the composition of the
media caused by compounds that are not analyzed can be
Using non-refined natural substrates such as lignocellulosic summarized under the commonly used term matrix effects.
material for the production of renewable fuels or chemi- Ideally, a data-driven algorithm, trained to monitor fermen-
cals has been central in the transition towards a sustainable tation processes, should measure the concentration changes
economy. The main operational challenges associated with of the analytes of interest (e.g., glucose, xylose and etha-
lignocellulosic material are the presence of mixed carbon nol for the case of cellulosic fermentations) independently
sources (including C6 and C5 sugars such as glucose and from concentration changes in other compounds or from
xylose, respectively), the presence of potent inhibitors matrix effects. Per se, purely data-driven models (such as
derived from the pretreatment of the biomass, and the high PLS regression) do not explicitly account for the dynam-
batch-to-batch variability of the substrates. Typically, cellu- ics of the fermentation, meaning that they will not be able
losic processes are run as fed-batch, where the concentration to differentiate the effect of two analytes with correlated
of substrates and inhibitors are kept below a certain level to dynamics independently. This situation raises a challenge
avoid inhibitions and to increase the space–time yield. Con- due to the high correlation between the dynamics of many
trol of the feed rate is a widely studied and often complex compounds in fermentation processes. That is, in the case
challenge in fermentation processes due to the non-linear of cellulose-to-ethanol fermentations, the uptake of glucose
nature of biological kinetics and to the limited ‘real-time’ or xylose is linearly correlated to the production of ethanol.
data extracted from such processes. Using lignocellulosic This situation is exemplified in Fig. 11a–d for a typical cellu-
feedstocks as a substrate for the fermentation increases this lose-to-ethanol fermentation. Decoupling such correlations
challenge due to its inherent natural variability. Not account- during the calibration (or training) of such models is cru-
ing for such variations often results in productivity losses cial to reach reliable predictions even when the dynamics of
and raises scheduling issues in the down-stream operations. the fermentation change. A common and efficient approach
Therefore, real-time monitoring methods are needed to consists of taking samples at different times during the fer-
implement advanced control schemes, whereas implement- mentation, and spiking them with different amounts of the
ing better controls inevitably leads to improved operations analytes of interest to attain partially uncorrelated samples
with quantifiable benefits [112]. [114] (Fig. 11e–h).
Systematic approaches to plan the calibration sets based
Open‑loop data‑driven monitoring of cellulose to ethanol on design of experiments (DoE) can successfully be imple-
fermentations mented in systems where the fermentation matrix can be
isolated from the rest of the analytes [115]. This allows to
When selecting a real-time monitoring scheme for fer- completely decouple the concentrations of the different ana-
mentation processes, it is crucial to account for the ease of lytes and to distribute the samples evenly along the design
implementation and for how the information can be used space.
to implement control strategies [113]. In cellulose-based While PLS models allow estimating the concentrations
fermentations, the commonly monitored variables such as of glucose, xylose and ethanol at each time point they have
pH or ­pO2 are easy to implement, but they do not deliver limited forecasting power and cannot be used to predict the
actionable information to develop control loops. Data-driven evolution of the fermentation. However, this information is
models are used to find correlations between the collected very valuable to determine the end-point of the fermenta-
spectra and the concentration of the different analytes of tion and to schedule the downstream operations. Cabaneros
interest. Linear models such as partial least squares (PLS) et al. [115] used a hybrid framework to incorporate the PLS
regressions are often used to model spectral data making use predictions of glucose into a mechanistic model of the fer-
of the linear correlation between the analyte concentration mentation to obtain high fidelity predictions of the progress
and the absorbance in the spectra defined by Lambert Beer’s of the fermentation.
law. Training such algorithms to monitor fermentation pro-
cesses efficiently is challenging due to the many sources of Closing the loop: feedback control of cellulose‑to‑ethanol
variation and the correlations between analyte concentra- fermentations
tions occurring during the fermentation. In cellulose-to-eth-
anol fermentations, S. cerevisiae mainly consumes glucose The objective of feed rate control in cellulose-to-ethanol
and xylose to produce ethanol, ­CO2, and biomass. However, fermentations is to keep the concentration of inhibitors or
to sustain its growth, yeast also takes up nitrogen or vitamins glucose at a certain set point to avoid inhibitions or catabo-
and produces other byproducts such as glycerol or acetate. lite repression. In a feedback control scheme the measured
Moreover, during the initial stage of the fermentation, S. variable (e.g., the concentration of inhibitors or substrate

13
Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964 959

A Glucsoe/Xylose Phase 1 B Glucsoe/Ethanol Phase 1 C Xylose/Ethanol Phase 1 D Xylose/Ethanol Phase 2

22 15 15 19
PCC = 0.988404 PCC = -0.999997 PCC = -0.988796 PCC = -1.000000

21.8
18.5

Ethanol (g/L)

Ethanol (g/L)

Ethanol (g/L)
Xylose (g/L)

21.6 10 10
18

21.4

17.5
21.2 5 5

17
21

20.8 0 0 16.5
0 10 20 30 40 0 10 20 30 40 20.8 21 21.2 21.4 21.6 21.8 22 10 12 14 16 18 20 22
Glucose (g/L) Glucose (g/L) Xylose (g/L) Xylose (g/L)

E Glucose/Xylose Spicked Phase 1 F Glucose/Ethanol Spicked Phase 1 G Xylose/Ethanol Spicked Phase 1 H Xylose/Ethanol Spicked Phase 2
32 20 20 28
PCC = 0.690868 PCC = -0.909719 PCC = -0.684441 PCC = -0.459849
31
18 18
26
30
16 16
Ethanol (g/L)

Ethanol (g/L)

Ethanol (g/L)
29 24
Xylose (g/L)

28 14 14
22
27 12 12

26 20
10 10
25
18
8 8
24

23 6 6 16
0 10 20 30 40 0 10 20 30 40 22 24 26 28 30 32 16 18 20 22 24 26 28
Glucose (g/L) Glucose (g/L) Xylose (g/L) Xylose (g/L)

Fig. 11  Pairwise correlation between the profiles of glucose, xylose, tions occurring during the glucose consuming phase. d and h Corre-
and ethanol in cellulose-to-ethanol batch fermentations in non-spiked lations during the xylose consuming phase
(a–d) and spiked samples (e–h). a–c and e–g correspond to correla-

predicted using data-driven models) serves as input to a to adjust the feed rate (Fig. 12). To implement robust real-
controller (such as a PID controller) to generate a signal time monitoring schemes for fermentation processes, it is
back to the manipulated variable (e.g., the controlled pump) fundamental to thoroughly understand the behavior of the

Fig. 12  Schematic representa- Holding tank Fermenter

tion of a feedback control loop
of the feed rate
Controlled
pump

In/on-line monitoring
equipment

Spectrum collection

Data-driven model

Present value of
controlled variable

Controller

Manipulated variable

13
960 Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964

monitoring method and the dynamics of the fermentation or processes based on off-line sampling and fixed process reci-
in any other unit operation as illustrated in [116, 117]. pes. These operational procedures, in many instances, ensure
that the products meet quality specifications but pay no atten-
tion to product and efficiency losses. The introduction of data-
Perspectives and barriers to implementation driven operations will require that the operators and chemists
adopt to dynamic schedules and on-line sampling methods,
The following perspectives and barriers to implementa- while embracing concepts, such as “real-time” product release,
tion are identified after analyzing in detail the case studies which are enabled by data-driven operations. This requires
presented together with the current state-of-the-art in bio- trust in data-driven operations by the operators, chemists and
process monitoring and operations. technical managers, and careful management of its imple-
mentation. The level of trust can be increased by improving
Sensing technologies the transparency of the data-driven operations and explicitly
disclosing the underlying assumptions and algorithms used.
The key to transitioning the current bio-based production
processes into data-driven operations is the ability to gather Value proposition
information-rich production data in real-time. This is the
primary objective of all three of the case studies presented With the current Industry 4.0 push that has been felt by the
and the focus of current efforts of the whole community. management level, there is increased demand for technolo-
Such efforts aim at the development of sensing technologies gies that fit these criteria. But, to achieve truly smart (bio)
and capabilities that go beyond the traditional temperature, manufacturing, noticeable fund allocations will be required
pressure and flow rate measurements, as they provide infor- from the corporate management that are more used to invest-
mation-rich process data that facilitate the data-driven oper- ing in “steel” for increased capacity (e.g. new equipment
ations of the future. However, these sensing technologies and/or new plants). These funds need to transition the cur-
(including the ones that are mentioned in the case studies) rent technologies in sensing, process monitoring and closed
require further development for them to be ready for long loop control to methods and technologies that are sufficiently
term use in industrial operations. Although the technologies robust for industrial operations. This also requires that the
have been proven on industrial trials and operations, their management organization at all levels is committed to make
technology readiness level (TRL) and maturity require fur- such investments. This further implies that the development
ther improvement before being robust for long term opera- and implementation cost of all these technologies must be
tions in an industrial setting. In addition, peripheral support- outweighed by the benefits in terms of efficiency, throughput
ing technologies are also under development that enable the and quality, as well as the environmental footprint.
use of these emerging technologies in an industrial bio-based
production environment. These range from chemometrics Regulation
and ML algorithms to GMP approved sampling ports and
sample transfer systems. There are strict regulatory hurdles for the implementa-
tion of new technologies in the bio-pharmaceutical and
Human element and organizational readiness food ingredients industries which compose a significant
part of the biomanufacturing industry. The current regula-
In comparison to the chemical industries, operators in the tions require operations to follow a strict recipe, whereas
biomanufacturing industry carry out a significant number the value proposition for these technologies is that they are
of complex time critical tasks. If data-driven concepts are to adapt the recipes based on “real time” information so
to improve these operations in the future, careful considera- as to allow for improved operations (throughout, quality,
tions must be given to the development of human machine efficiency, resource usage). This warrants a change in regu-
interfaces (HMI’s) that effectively communicate the out- lations following a QbD approach which has already been
comes of these data-driven methods. In addition, such meth- proposed, where a “region” of process operations is allowed
ods must pay heed to the human limitations when suggesting as opposed to strict recipes.
operational actions. For example, if a continuous addition to
a process over an extended period is considered optimal, this
turns out to be an infeasible task for an operator. Rather, this Conclusion
operational action might need to be modified to one addition
to the process at an optimal time period. The promise of improved efficiency brought by smart
From an organizational point of view, the operators and biomanufacturing has sparked the interest of both aca-
chemists are familiar with operating complex bio-based demia and industry towards the development of smart

13
Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964 961

technologies. However, the actual implementation of these as an activating teaching tool. Educ Chem Eng 30:20–31. https​
technologies in an industrial setting will require significant ://doi.org/10.1016/j.ece.2019.09.002
3. Iriondo R (2020) Machine learning vs. AI, important differences
further efforts. As illustrated by the first two case stud- between them. Medium. https​://mediu​m.com/towar​ds-artif​i cial​
ies, there are still major technical challenges that must be -intel​ligen​ce/diffe​rence​s-betwe​en-ai-and-machi​ne-learn​ing-and-
overcome to ensure that sufficiently large data sets can be why-it-matte​rs-1255b​182fc​6. Accessed 9 May 2020
collected. These two case studies also highlight the ability 4. Markarian J (2018) Modernizing pharma manufacturing. Pharm
Technol 42(4):20–25
to adopt already developed ML algorithms so as to accel- 5. FDA (2004) Guidance for industry, PAT-A framework for inno-
erate the implementation of specific process monitoring vative pharmaceutical development, manufacturing and quality
needs. The last case study demonstrates how fully auto- assurance
mated closed loop control can be achieved in a fermenta- 6. Randek J, Mandenius CF (2018) On-line soft sensing in upstream
bioprocessing. Crit Rev Biotechnol 38(1):106–121. https​://doi.
tion operation by leveraging large amounts of information- org/10.1080/07388​551.2017.13122​71
rich process data. As detailed in the perspectives section, 7. Challener CA (2014) Improving PAT for biologics with online
there are significant barriers that must also be addressed spectroscopy and multivariate data analysis. Biopharm Int
alongside the development of the core sensing, monitoring 27(12):18–21
8. Claßen J, Aupert F, Reardon KF, Solle D, Scheper T (2017)
and control technologies. All in all, this manuscript illus- Spectroscopic sensors for in-line bioprocess monitoring in
trates that there are exciting ongoing developments in the research and pharmaceutical industrial application. Anal Bio-
sensor area for monitoring and control of bio-based pro- anal Chem 409(3):651–666. https​: //doi.org/10.1007/s0021​
cesses that hold the promise of moving the whole industry 6-016-0068-x
9. Mears L, Stocks SM, Albaek MO, Sin G, Gernaey KV (2017)
towards smart biomanufacturing in the future. Mechanistic fermentation models for process design, monitor-
ing, and control. Trends Biotechnol 35(10):914–924. https​://doi.
Acknowledgements The authors wish to acknowledge the financial org/10.1016/j.tibte​ch.2017.07.002
support provided by the following grants: (i) the Danish Council for 10. Gernaey KV, Lantz AE, Tufvesson P, Woodley JM, Sin G
Independent Research in the frame of the DFF FTP research project (2010) Application of mechanistic models to fermentation and
GREENLOGIC (Grant agreement number 7017-00175A) and by the biocatalysis for next-generation processes. Trends Biotechnol
Novo Nordisk Foundation in the frame of the Fermentation-Based 28(7):346–354. https​://doi.org/10.1016/j.tibte​ch.2010.03.006
Biomanufacturing education initiative (Grant agreement number 11. Veloso AC, Ferreira EC (2017) Online analysis for industrial
NNF17SA0031362); (ii) the Energy Technology Development and bioprocesses: broth analysis. Curr Develop Biotechnol Bioeng
Demonstration Program (EUDP) in the frame of the project “Demon- Bioprocess Bioreact Controls. https​://doi.org/10.1016/B978-0-
stration of 2G ethanol production in full scale” (Grant number 64015- 444-63663​-8.00023​-9
0642); (iii) Greater Copenhagen Food Innovation project (CPH Food), 12. Lourenço ND, Lopes JA, Almeida CF, Sarraguça MC, Pinheiro
Novozymes, from EUs regional fund (BIOPRO-SMV project) and from HM (2012) Bioreactor monitoring with spectroscopy and chemo-
Innovation Fund Denmark through the BIOPRO2 strategic research metrics: a review. Anal Bioanal Chem 404(4):1211–1237. https​
center (Grant number 4105-00020B); and, (iv) the Novo Nordisk Foun- ://doi.org/10.1007/s0021​6-012-6073-9
dation in the frame of the ‘Accelerated Innovation in Manufacturing 13. Alford JS (2006) Bioprocess control: advances and chal-
Biologics’ (AIMBio) project (Grant number NNF19SA0035474). lenges. Comput Chem Eng 30(10–12):1464–1475. https​://doi.
org/10.1016/j.compc​hemen​g.2006.05.039
Open Access This article is licensed under a Creative Commons Attri- 14. Arnold SA, Harvey LM, McNeil B, Hall JW (2003) Employing
bution 4.0 International License, which permits use, sharing, adapta- near-infrared spectroscopic methods of analysis for fermenta-
tion, distribution and reproduction in any medium or format, as long tion monitoring and control. Part 2: Implementation strategies.
as you give appropriate credit to the original author(s) and the source, Biopharm Int 16:47–70
provide a link to the Creative Commons licence, and indicate if changes 15. Arnold SA, Harvey LM, McNeil B, Hall JW (2002) Employing
were made. The images or other third party material in this article are near-infrared spectroscopic methods of analysis for fermentation
included in the article’s Creative Commons licence, unless indicated monitoring and control. Part 1: method development. Biopharm
otherwise in a credit line to the material. If material is not included in Int 15:26–34
the article’s Creative Commons licence and your intended use is not 16. Helmel M et al (2015) Intact cell mass spectrometry as a pro-
permitted by statutory regulation or exceeds the permitted use, you will gress tracking tool for batch and fed-batch fermentation pro-
need to obtain permission directly from the copyright holder. To view a cesses. Anal Biochem 470:25–33. https​://doi.org/10.1016/j.
copy of this licence, visit http://creat​iveco​mmons​.org/licen​ses/by/4.0/. ab.2014.10.008
17. Slouka C et al (2017) Low-frequency electrochemical imped-
ance spectroscopy as a monitoring tool for yeast growth in
industrial brewing processes. Chemosensors 5(3):24. https​://
doi.org/10.3390/chemo​senso​rs503​0024
References 18. Chadwick DT, McDonnell KP, Brennan LP, Fagan CC, Everard
CD (2014) Evaluation of infrared techniques for the assessment
1. Eifert T, Eisen K, Maiwald M, Herwig C (2020) Current and of biomass and biofuel quality parameters and conversion tech-
future requirements to industrial analytical infrastructure—part nology processes: a review. Renew Sustain Energy Rev 30:672–
2: smart sensors. Anal Bioanal Chem 412(9):2037–2045. https​ 681. https​://doi.org/10.1016/j.rser.2013.11.006
://doi.org/10.1007/s0021​6-020-02421​-1 19. Correia DM et al (2014) Dietary sugars analysis: quantification of
2. Udugama IA, Gernaey KV, Taube MA, Bayer C (2020) A novel fructooligossacharides during fermentation by HPLC-RI method.
use for an old problem: the Tennessee Eastman challenge process Front Nutr. https​://doi.org/10.3389/fnut.2014.00011​

13
962 Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964

20. Polizzi KM, Kontoravdi C (2015) Genetically-encoded bio- Biores Technol 134:143–150. https​://doi.org/10.1016/j.biort​
sensors for monitoring cellular stress in bioprocessing. Curr ech.2013.01.136
Opin Biotechnol 31:50–56. https​: //doi.org/10.1016/j.copbi​ 35. Kim S et al (2015) Real-time monitoring of glycerol and
o.2014.07.011 methanol to enhance antibody production in industrial Pichia
21. Schenkmayerová A et al (2015) Whole-cell Gluconobacter pastoris bioprocesses. Biochem Eng J 94:115–124. https​://doi.
oxydans biosensor for 2-phenylethanol biooxidation monitor- org/10.1016/j.bej.2014.12.002
ing. Anal Chim Acta 854:140–144. https​://doi.org/10.1016/j. 36. Kruid J, Fogel R, Limson J (2014) Voltammetric investigation of
aca.2014.11.012 complex growth media at a bare glassy carbon electrode: a case
22. Bäcker M, Rakowski D, Poghossian A, Biselli M, Wagner P, study of oxytetracycline. Electrochim Acta 128:41–47. https​://
Schöning MJ (2013) Chip-based amperometric enzyme sensor doi.org/10.1016/j.elect​acta.2013.08.188
system for monitoring of bioprocesses by flow-injection analy- 37. Oh S-K, Yoo SJ, Jeong DH, Lee JM (2013) Real-time estima-
sis. J Biotechnol 163(4):371–376. https​://doi.org/10.1016/j.jbiot​ tion of glucose concentration in algae cultivation system using
ec.2012.03.014 Raman spectroscopy. Biores Technol 142:131–137. https​://doi.
23. Dietzsch C, Spadiut O, Herwig C (2013) On-line multiple com- org/10.1016/j.biort​ech.2013.05.008
ponent analysis for efficient quantitative bioprocess develop- 38. Picard A, Daniel I, Montagnac G, Oger P (2007) In situ moni-
ment. J Biotechnol 163(4):362–370. https​://doi.org/10.1016/j. toring by quantitative Raman spectroscopy of alcoholic fer-
jbiot​ec.2012.03.010 mentation by Saccharomyces cerevisiae under high pressure.
24. Beuermann T, Egly D, Geoerg D, Klug KI, Storhas W, Methner Extremophiles 11(3):445–452. https​://doi.org/10.1007/s0079​
F-J (2012) On-line carbon balance of yeast fermentations using 2-006-0054-x
miniaturized optical sensors. J Biosci Bioeng 113(3):399–405. 39. Quintelas C, Mesquita DP, Lopes JA, Ferreira EC, Sousa C
https​://doi.org/10.1016/j.jbios​c.2011.10.016 (2015) Near-infrared spectroscopy for the detection and quanti-
25. Abad S, Pérez X, Planas A, Turon X (2014) Rapid monitoring of fication of bacterial contaminations in pharmaceutical products.
glycerol in fermentation growth media: facilitating crude glyc- Int J Pharm 492(1–2):199–206. https​://doi.org/10.1016/j.ijpha​
erol bioprocess development. Talanta 121:210–214. https​://doi. rm.2015.07.005
org/10.1016/j.talan​ta.2013.12.022 40. Sampaio PN, Sales KC, Rosa FO, Lopes MB, Calado CR (2014)
26. Alves-Rausch J, Bienert R, Grimm C, Bergmaier D (2014) Real In situ near infrared spectroscopy monitoring of cyprosin produc-
time in-line monitoring of large scale Bacillus fermentations with tion by recombinant Saccharomyces cerevisiae strains. J Biotech-
near-infrared spectroscopy. J Biotechnol 189:120–128. https​:// nol 188:148–157. https​://doi.org/10.1016/j.jbiot​ec.2014.07.454
doi.org/10.1016/j.jbiot​ec.2014.09.004 41. Tiwari S, Suraishkumar GK, Chandavarkar A (2013) Robust
27. Cruz MV, Sarraguça MC, Freitas F, Lopes JA, Reis MAM (2015) near-infra-red spectroscopic probe for dynamic monitoring of
Online monitoring of P(3HB) produced from used cooking oil critical nutrient ratio in microbial fermentation processes. Bio-
with near-infrared spectroscopy. J Biotechnol 194:1–9. https​:// chem Eng J 71:47–56. https​://doi.org/10.1016/j.bej.2012.11.009
doi.org/10.1016/j.jbiot​ec.2014.11.022 42. Veloso ACA, Rodrigues LR, Dias LG, Peres AM (2012) UV
28. Dias LG et al (2009) UV spectrophotometry method for the spectrophotometry method for dietary sugars. In: Preedy V (ed)
monitoring of galacto-oligosaccharides production. Food Chem Food and nutritional components in focus, vol 14. Royal Society
113(1):246–252. https:​ //doi.org/10.1016/j.foodch​ em.2008.06.072 of Chemistry, Coimbra, pp 229–248
29. Fuentes-Grünewald C, Bayliss C, Zanain M, Pooley C, Scola- 43. Prasad K, Ranjan RK, Lutfi Z, Pandey H (2009) Biosensors:
macchia M, Silkina A (2015) Evaluation of batch and semi-con- applications and overview in industrial automation. Int J Appl
tinuous culture of Porphyridium purpureum in a photobioreactor Bioeng 3(1):66–70. https​://doi.org/10.18000​/ijabe​g.10041​
in high latitudes using Fourier Transform Infrared spectroscopy 44. Borisov SM, Wolfbeis OS (2008) Optical biosensors. Chem Rev
for monitoring biomass composition and metabolites produc- 108(2):423–461. https​://doi.org/10.1021/cr068​105t
tion. Biores Technol 189:357–363. https:​ //doi.org/10.1016/j.biort​ 45. Grieshaber D, MacKenzie R, Vörös J, Reimhult E (2008) Elec-
ech.2015.04.042 trochemical biosensors—sensor principles and architectures.
30. Grassi S, Amigo JM, Lyndgaard CB, Foschino R, Casiraghi E Sensors. https​://doi.org/10.3390/s8031​4000
(2014) Assessment of the sugars and ethanol development in 46. Yakovleva M, Bhand S, Danielsson B (2013) The enzyme ther-
beer fermentation with FT-IR and multivariate curve resolution mistor—a realistic biosensor concept. A critical review. Anal
models. Food Res Int 62:602–608. https:​ //doi.org/10.1016/j.foodr​ Chim Acta 766:1–12. https​://doi.org/10.1016/j.aca.2012.12.004
es.2014.03.058 47. Pohanka M (2018) Overview of piezoelectric biosensors, immu-
31. Grote B, Zense T, Hitzmann B (2014) 2D-fluorescence and mul- nosensors and DNA sensors and their applications. Materials
tivariate data analysis for monitoring of sourdough fermentation (Basel, Switzerland) 11(3):448. https​://doi.org/10.3390/ma110​
process. Food Control 38:8–18. https​://doi.org/10.1016/j.foodc​ 30448​
ont.2013.09.039 48. Rhee JII, Ritzka A, Scheper T (2004) On-line monitoring and
32. Hakemeyer C, Strauss U, Werz S, Jose GE, Folque F, Menezes control of substrate concentrations in biological processes by
JC (2012) At-line NIR spectroscopy as effective PAT monitoring flow injection analysis systems. Biotechnol Bioprocess Eng
technique in Mab cultivations during process development and 9(3):156–165. https​://doi.org/10.1007/BF029​42286​
manufacturing. Talanta 90:12–21. https:​ //doi.org/10.1016/j.talan​ 49. Pilas J, Selmer T, Keusgen M, Schöning MJ (2019) Screen-
ta.2011.12.042 printed carbon electrodes modified with graphene oxide for the
33. Iversen JA, Berg RW, Ahring BK (2014) Quantitative monitor- design of a reagent-free NAD+-dependent biosensor array. Anal
ing of yeast fermentation using Raman spectroscopy. Anal Bio- Chem 91(23):15293–15299. https​://doi.org/10.1021/acs.analc​
anal Chem 406(20):4911–4919. https​://doi.org/10.1007/s0021​ hem.9b044​81
6-014-7897-2 50. Semenova D et al (2018) Mechanistic modeling of cyclic vol-
34. Kedia G, Passanha P, Dinsdale RM, Guwy AJ, Lee M, Esteves tammetry: a helpful tool for understanding biosensor principles
SR (2013) Addressing the challenge of optimum polyhydroxy- and supporting design optimization. Sens Actuators B Chem
alkanoate harvesting: monitoring real time process kinetics 259:945–955. https​://doi.org/10.1016/J.SNB.2017.12.088
and biopolymer accumulation using dielectric spectroscopy.

13
Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964 963

51. Biechele P, Busse C, Solle D, Scheper T, Reardon K (2015) Sen- Biosyst Eng 32(3):381–388. https​: //doi.org/10.1007/s0044​
sor systems for bioprocess monitoring. Eng Life Sci 15(5):469– 9-008-0257-x
488. https​://doi.org/10.1002/elsc.20150​0014 69. Sharma S, Tambe SS (2014) Soft-sensor development for bio-
52. Becker et al (2006) Future aspects of bioprocess monitor- chemical systems using genetic programming. Biochem Eng J
ing. Adv Biochem Eng Biotechnol 105:249–293. https​://doi. 85:89–100. https​://doi.org/10.1016/j.bej.2014.02.007
org/10.1007/10_2006_036 70. Krause D, Hussein MA, Becker T (2015) Online monitoring of
53. Samorski M, Mueller-Newen G, Buechs J (2005) Quasi-continu- bioprocesses via multivariate sensor prediction within swarm
ous combined scattered light and fluorescence measurements: a intelligence decision making. Chemometr Intell Lab Syst
novel measurement technique for shaken microtiter plates. Bio- 145:48–59. https​://doi.org/10.1016/j.chemo​lab.2015.04.012
technol Bioeng 92(1):61–68 71. Abu-Absi NR, Martel RP, Lanza AM, Clements SJ, Borys MC,
54. Kensy F, Zang E, Faulhammer C, Tan R-K, Büchs J (2009) Li ZJ (2014) Application of spectroscopic methods for moni-
Validation of a high-throughput fermentation system based on toring of bioprocesses and the implications for the manufac-
online monitoring of biomass and fluorescence in continuously ture of biologics. Pharm Bioprocess 2(3):267–284. https​://doi.
shaken microtiter plates. Microb Cell Fact 8(1):31. https​://doi. org/10.4155/pbp.14.24
org/10.1186/1475-2859-8-31 72. Sundström H, Enfors S-O (2008) Software sensors for fermenta-
55. Rolinger L, Rüdt M, Hubbuch J (2020) A critical review of recent tion processes. Bioprocess Biosyst Eng 31(2):145–152. https​://
trends, and a future perspective of optical spectroscopy as PAT in doi.org/10.1007/s0044​9-007-0157-5
biopharmaceutical downstream processing. Anal Bioanal Chem 73. Golabgir A, Hoch T, Zhariy M, Herwig C (2015) Observability
412(9):2047–2064. https:​ //doi.org/10.1007/s00216​ -020-02407-​ z analysis of biochemical process models as a valuable tool for
56. Analytictrust increase the value of your analytic instruments. the development of mechanistic soft sensors. Biotechnol Prog
AnalyticTrust, 2020. www.analy​tictr​ust.com. August 2020 31(6):1703–1715. https​://doi.org/10.1002/btpr.2176
57. Esmonde-White KA, Cuellar M, Uerpmann C, Lenain B, Lewis 74. Ehgartner D, Sagmeister P, Herwig C, Wechselberger P (2015)
IR (2017) Raman spectroscopy as a process analytical technol- A novel real-time method to estimate volumetric mass biodensity
ogy for pharmaceutical manufacturing and bioprocessing. Anal based on the combination of dielectric spectroscopy and soft-
Bioanal Chem 409(3):637–649. https​://doi.org/10.1007/s0021​ sensors. J Chem Technol Biotechnol 90(2):262–272. https​://doi.
6-016-9824-1 org/10.1002/jctb.4469
58. Glassey J (2013) Multivariate data analysis for advancing the 75. Ödman P, Johansen CL, Olsson L, Gernaey KV, Lantz AE (2009)
interpretation of bioprocess measurement and monitoring On-line estimation of biomass, glucose and ethanol in Saccha-
data. Adv Biochem Eng Biotechnol 132:167–191. https​://doi. romyces cerevisiae cultivations using in-situ multi-wavelength
org/10.1007/10_2012_171 fluorescence and software sensors. J Biotechnol 144(2):102–112.
59. Pomerantsev AL, Ye O (2012) Rodionova, “Process analytical https​://doi.org/10.1016/j.jbiot​ec.2009.08.018
technology: a critical view of the chemometricians”. J Chemom 76. Melcher M et al (2015) The potential of random forest and
26(6):299–310. https​://doi.org/10.1002/cem.2445 neural networks for biomass and recombinant protein mod-
60. Henriques JG, Buziol S, Stocker E, Voogd A, Menezes JC (2010) eling in Escherichia coli fed-batch fermentations. Biotechnol J
Monitoring mammalian cell cultivations for monoclonal anti- 10(11):1770–1782. https​://doi.org/10.1002/biot.20140​0790
body production using near-infrared spectroscopy. Adv Biochem 77. von Stosch M et al (2014) Hybrid modeling for quality by
Eng Biotechnol. https​://doi.org/10.1007/10_2009_11 design and PAT-benefits and challenges of applications in biop-
61. Navrátil M, Norberg A, Lembrén L, Mandenius C-F (2005) On- harmaceutical industry. Biotechnol J 9(6):719–726. https​://doi.
line multi-analyzer monitoring of biomass, glucose and acetate org/10.1002/biot.20130​0385
for growth rate control of a Vibrio cholerae fed-batch cultiva- 78. Galvanauskas V, Simutis R, Lübbert A (2004) Hybrid process
tion. J Biotechnol 115(1):67–79. https:​ //doi.org/10.1016/j.jbiot​ models for process optimisation, monitoring and control. Bio-
ec.2004.07.013 process Biosyst Eng 26(6):393–400. https​://doi.org/10.1007/
62. Roggo Y, Chalus P, Maurer L, Lema-Martinez C, Edmond A, s0044​9-004-0385-x
Jent N (2007) A review of near infrared spectroscopy and chemo- 79. Teixeira A et al (2005) Modelling and optimization of a recom-
metrics in pharmaceutical technologies. J Pharm Biomed Anal binant BHK-21 cultivation process using hybrid grey-box sys-
44(3):683–700. https​://doi.org/10.1016/j.jpba.2007.03.023 tems. J Biotechnol 118(3):290–303. https​://doi.org/10.1016/j.
63. Cervera AE, Petersen N, Lantz AE, Larsen A, Gernaey KV jbiot​ec.2005.04.024
(2009) Application of near-infrared spectroscopy for monitoring 80. Ohadi K, Legge RL, Budman HM (2015) Development of a
and control of cell culture and fermentation. Biotechnol Progr. soft-sensor based on multi-wavelength fluorescence spectros-
https​://doi.org/10.1002/btpr.280 copy and a dynamic metabolic model for monitoring mamma-
64. Busse C, Biechele P, de Vries I, Reardon KF, Solle D, Scheper lian cell cultures. Biotechnol Bioeng 112(1):197–208. https:​ //doi.
T (2017) Sensors for disposable bioreactors. Eng Life Sci org/10.1002/bit.25339​
17(8):940–952. https​://doi.org/10.1002/elsc.20170​0049 81. Golabgir A, Herwig C (2016) Combining mechanistic modeling
65. Zimmermann R, Fiabane L, Gasteuil Y, Volk R, Pinton J-F (2013) and Raman Spectroscopy for real-time monitoring of fed-batch
Measuring Lagrangian accelerations using an instrumented penicillin production. Chem Ing Tec 88(6):764–776. https​://doi.
particle. Phys Scr T155:14063. https​://doi.org/10.1088/0031- org/10.1002/cite.20150​0101
8949/2013/t155/01406​3 82. Selinummi J et al (2009) Bright field microscopy as an alternative
66. Luttmann R et al (2012) Soft sensors in bioprocessing: a status to whole cell fluorescence in automated analysis of macrophage
report and recommendations. Biotechnol J 7(8):1040–1048. https​ images. PLoS ONE 4(10):e7497. https​://doi.org/10.1371/journ​
://doi.org/10.1002/biot.20110​0506 al.pone.00074​97
67. Assis A, Marciel R (2000) Soft sensors development for 83. Ettinger A, Wittmann T (2014) Fluorescence live cell imaging.
on-line bioreactor state estimation. Comput Chem Eng Methods Cell Biol 123:77–94
24(2–7):1099–1103 84. Rieder CL (2003) Mitosis through the microscope: advances in
68. Veloso ACA, Rocha I, Ferreira EC (2009) Monitoring of fed- seeing inside live dividing cells. Science 300(5616):91–96. https​
batch E. coli fermentations with software sensors. Bioprocess ://doi.org/10.1126/scien​ce.10821​77

13
964 Journal of Industrial Microbiology & Biotechnology (2020) 47:947–964

85. Telser A (2002) Fundamentals of light microscopy and electronic 103. View particles in real time—ensure comprehensive understand-
imaging. Shock 17(5):442. https​://doi.org/10.1097/00024​382- ing (2019). Syntech Innovation, 2015. www.synte​chinn​ovati​
20020​5000-00022​ on.com. August 2020
86. Mondal PP, Diaspro A (2014) Fundamentals of fluorescence 104. Microscope probes - high resolution and magnification. Sopat
microscopy. Springer Netherlands, Dordrecht (2020). sopat​.de/en/micro​scope​-probe​s. August 2020
87. Zernike F (1942) Phase contrast, a new method for the micro- 105. Lasentec (1991) Operations manual for PAR-TEC 100. Lasentec
scopic observation of transparent objects part II. Physica 106. Particletech solution (2019) ParticleTech. https:​ //partic​ letec​ h.dk/
9(10):974–986. https:​ //doi.org/10.1016/S0031-​ 8914(42)80079-​ 8 parti​clete​chsol​ution​/
88. Smith CL (2008) Basic confocal microscopy. In: Current pro- 107. Nielsen RF, Kermani NA, la Cour Freiesleben L, Gernaey KV,
tocols in neuroscience 2(56):2.2. https​://doi.org/10.1002/04711​ Mansouri SS (2019) Novel strategies for predictive particle
42301​.ns020​2s56 monitoring and control using advanced image analysis. In: 29th
89. Frigault MM, Lacoste J, Swift JL, Brown CM (2009) Live-cell European symposium on computer aided process engineering,
microscopy—tips and tools. J Cell Sci 122(6):753–767. https​:// vol 46, pp 1435–1440. https:​ //doi.org/10.1016/B978-0-12-81863​
doi.org/10.1242/jcs.03383​7 4-3.50240​-X
90. Walter RJ (1981) Computer-enhanced video microscopy: digi- 108. Nielsen RF, Nazemzadeh N, Sillesen LW, Andersson MP, Ger-
tally processed microscope images can be produced in real time. naey KV, Mansouri SS (2020) Hybrid machine learning assisted
In: Proceedings of the National Academy of Sciences of the modelling framework for particle processes. Compu Chem Eng.
United States of America, vol 78, no. 11 https​://doi.org/10.1016/j.compc​hemen​g.2020.10691​6
91. Coutu DL, Schroeder T (2013) Probing cellular processes by 109. Nazemzadeh N et al (2020) Integration of computational chem-
long-term live imaging—historic problems and current solutions. istry and artificial intelligence for multi-scale modeling of bio-
J Cell Sci 126(17):3805–3815. https:​ //doi.org/10.1242/jcs.11834​ processes (accepted). In: 30th European symposium on computer
9 aided process engineering
92. Fredborg M, Rosenvinge FS, Spillum E, Kroghsbo S, Wang M, 110. Govindan K, Jafarian A, Khodaverdi R, Devika K (2014) Two-
Sondergaard TE (2015) Rapid antimicrobial susceptibility testing echelon multiple-vehicle location-routing problem with time
of clinical isolates by digital time-lapse microscopy. Eur J Clin windows for optimization of sustainable supply chain net-
Microbiol Infect Dis 34(12):2385–2394. https:​ //doi.org/10.1007/ work of perishable food. Int J Prod Econ 152:9–28. https​://doi.
s1009​6-015-2492-9 org/10.1016/j.ijpe.2013.12.028
93. Uggerhøj LE et al (2015) Rational design of alpha-helical antimi- 111. Abadi M, Agarwal A, Barham P, Brevdo E, Chen Z, Citro C, Cor-
crobial peptides: do’s and don’ts. ChemBioChem 16(2):242–253. rado GS et al (2016) TensorFlow: large-scale machine learning
https​://doi.org/10.1002/cbic.20140​2581 on heterogeneous distributed systems. arXiv​:1603.04467​
94. Kjeldsen T, Sommer M, Olsen JE (2015) Extended spectrum 112. Udugama IA et al (2018) A systematic methodology for compre-
β-lactamase-producing Escherichia coli forms filaments as hensive economic assessment of process control structures. Ind
an initial response to cefotaxime treatment. BMC Microbiol Eng Chem Res 57(39):13116–13130. https:​ //doi.org/10.1021/acs.
15(1):63. https​://doi.org/10.1186/s1286​6-015-0399-3 iecr.8b018​83
95. Ashley N, Jones M, Ouaret D, Wilding J, Bodmer WF (2014) 113. Cabaneros P, Feldman H, Mauricio-iglesias M, Junicke H,
Rapidly derived colorectal cancer cultures recapitulate parental Kjøbsted J, Gernaey KV (2019) Benchmarking real-time moni-
cancer characteristics and enable personalized therapeutic assays. toring strategies for ethanol production from lignocellulosic
J Pathol 234(1):34–45. https​://doi.org/10.1002/path.4371 biomass. Biomass Bioenerg 127:105296–105296. https​://doi.
96. Fredborg M et al (2013) Real-time optical antimicrobial suscep- org/10.1016/j.biomb​ioe.2019.10529​6
tibility testing. J Clin Microbiol 51(7):2047–2053. https​://doi. 114. Pontius K, Junicke H, Gernaey KV, Bevilacqua M (2020) Moni-
org/10.1128/JCM.00440​-13 toring yeast fermentations by nonlinear infrared technology and
97. Fredborg M, Rosenvinge FS, Spillum E, Kroghsbo S, Wang M, chemometrics—understanding process correlations and indirect
Sondergaard TE (2015) Automated image analysis for quantifica- predictions. Appl Microbiol Biotechnol. https​://doi.org/10.1007/
tion of filamentous bacteria. BMC Microbiol 15(1):255. https​:// s0025​3-020-10604​-0
doi.org/10.1186/s1286​6-015-0583-5 115. Cabaneros Lopez P et al (2020) Towards a Digital Twin: a hybrid
98. Tibayrenc P, Preziosi-Belloy L, Roger JM, Ghommidh C data-driven and mechanistic digital shadow to forecast the evolu-
(2010) Assessing yeast viability from cell size measurements? tion of lignocellulosic fermentations. Biofuels Bioprod Biorefin.
J Biotechnol 149(1–2):74–80. https​://doi.org/10.1016/j.jbiot​ https​://doi.org/10.1002/bbb.2108
ec.2010.06.019 116. Udugama IA, Wolfenstetter F, Kirkpatrick R, Yu W, Young
99. Camisard V, Brienne JP, Baussart H, Hammann J, Suhr H (2002) BR (2017) A comparison of a novel robust decentralised con-
Inline characterization of cell concentration and cell volume in trol strategy and MPC for industrial high purity, high recovery,
agitated bioreactors using in situ microscopy: application to multicomponent distillation. ISA Trans 69:222–233. https​://doi.
volume variation induced by osmotic stress. Biotechnol Bioeng. org/10.1016/j.isatr​a.2017.04.008
https​://doi.org/10.1002/bit.10178​ 117. Udugama IA, Munir MT, Kirkpatrick R, Young BR, Yu W
100. Tyson CB, Lord PG, Wheals AE (1979) Dependency of size (2018) Side draw control design for a high purity multi-com-
of Saccharomyces cerevisiae cells on growth rate. J Bacteriol ponent distillation column. ISA Trans 76:167–177. https​://doi.
138(1):92–98 org/10.1016/j.isatr​a.2018.03.006
101. Marbà-Ardébol A-M, Emmerich J, Neubauer P, Junne S (2017) 118. Udugama IA, Gargalo CL, Yamashita Y, Taube MA, Palazoglu
Single-cell-based monitoring of fatty acid accumulation in A, Young BR, Gernaey KV, Kulahci M, Bayer C (2020) The role
Crypthecodinium cohnii with three-dimensional holographic of big data in industrial (bio)chemical process operations. Ind
and in situ microscopy. Process Biochem 52:223–232. https​:// Eng Chem Rese 59(34):15283–15297
doi.org/10.1016/j.procb​io.2016.11.003
102. Pontius K (2019) Monitoring of bioprocesses opportunities and Publisher’s Note Springer Nature remains neutral with regard to
challenges monitoring of bioprocesses opportunities and chal- jurisdictional claims in published maps and institutional affiliations.
lenges. PhD Thesis, Technical University of Denmark, Denmark

13