Joseph S.

Alford
Consultant

The use of complex living organisms, process variation,

and lack of real-time measurements of key parameters
are some of the challenges involved in automating a
bioprocess. This article explains an indirect approach to
monitoring and controlling a fermentation process.

he automation of chemical processes began in the

continuous petrochemical industries in the mid-20th
century. University process-control courses and commercial automation products focused primarily on continuous linear processes running under steady-state conditions.
In the past few decades, batch and discrete manufacturing operations have started to benefit from automatic
control. As a result, both vendor software and undergraduate courses have been challenged to integrate the unique
requirements of these types of processes.
Most commercial bioprocesses are batch operations
that are not continuous, linear, or steady-state. Automating
them is significantly more complex than implementing a
few proportional-integral-derivative (PID) controllers.
This article examines some of the unique characteristics
of bioprocesses, as well as a few selected measurement,
modeling, and control techniques that are useful in automating them.

What is a bioprocess?
A bioprocess is a process in which the desired product (or
its precursor) for instance, antibiotics such as penicillin or
proteins such as insulin is produced by living biological
material (e.g., fungal cells, E. coli bacteria, yeast, or mammalian cells) in a bioreactor (with various supporting unit
operations). Most bioprocesses consist of three or four general phases that are conducted as a sequence of batch steps:
1. growing the living biological material
2. using the living biological material to make a desired
product
44

www.aiche.org/cep

November 2009 CEP

3. separating the desired product from process materials

and then purifying it
4. in some cases, modifying the product via chemical
reaction before final purification.
Given the large diversity of biological processes, there
are exceptions to the above paradigm. Sometimes the cell
mass itself is the desired product, such as when it is used as
a protein food supplement. In other cases, it is the activity
of living cells that is important, such as the degradation of
matter in waste-treatment systems.
Figure 1 shows some of the equipment involved in typical commercial bioprocesses. Note the use of unit operations such as fermentation and chromatography separations
that are not commonly found in continuous processing
industries. Figure 2 provides more detail on the components of a bioreactor, which is one of the primary unit
operations of most industrial bioprocesses and the main
focus of this article.
The cycle of activities for a bioreactor generally lasts
from one day to one or more weeks, and typically includes
the following steps:
1. cleaning
2. adding initial ingredients
3. sterilizing
4. inoculation with cells (from the seed tank)
5. cell growth
6. product production
7. harvesting.
Most of the steps, including cleaning, can be automated.
The cell growth and product production steps require

Seed Tank

Fermentor

Centrifuge

Chromatography
Column

Chemical
Reactor

Chromatography
Column

Crystallizer

S Figure 1. Bioprocesses employ some unit operations that are unlike those of a traditional continuous chemical-manufacturing process.

maintaining an appropriate environment for the cells, which

usually includes the automated feeding of various nutrients
(e.g., substrates) and the removal of undesired waste products (e.g., CO2).

Most bioprocesses are run in batch mode

Although batch processes are typically more labor
intensive and operate at lower overall productivities than
continuous processes, most bioprocesses are run in batch
mode for several reasons:
criteria for forward processing. Most commercial
bioprocesses in regulated industries (e.g., pharmaceutical)
include defined attributes known as criteria for forward
processing. These are typically attributes related to the
quality of the active pharmaceutical ingredient (API) or the
final product, and determining compliance with them often
requires laboratory assays that may take hours or days to
complete. It can be appropriate to stop and hold a process
after the completion of certain steps until it is known that
the criteria for forward processing have been satisfied.
cells finite life. Living organisms have a finite life.
Cells can usually survive for many generations (i.e., cell
divisions) in the favorable environment of a well-controlled
bioreactor, but ultimately they will slow down, die and lyse
(break apart).
product accumulation within cells. Some fermentations make a product that is retained within the cell walls,
rather than being secreted. Such cells can only hold a finite
amount of product. When the limit is approached, it is time
to stop the fermentation.
product degradation. Some products slowly degrade
in the bioreactor. In such cases, the fermentation should
be stopped when the rate of product degradation becomes
significant compared to the rate of product synthesis.
cell mutation. When cells divide, there is a small
but finite chance that a mutation will occur, such that the
daughter cells are not identical to the parent cell. Most
mutant cells will yield less product and, theoretically, could

Agitator
Samples

Feeds
(Nutrients,
Antifoam, Acid/Base)

Sterile Inlet Gases

(e.g., Air, O2, CO2)

Exhaust
Gases Out

To
Process Mass
Spectrometer

Broth
Steam or
Cooling Water Out

Sparger
Agitator
Blade
Baffle

pH
Temperature
Steam or
Cooling Water In

Dissolved
Oxygen

S Figure 2. The bioreactor, or fermentor, is the heart of a bioprocess.

produce higher levels of undesired metabolites. Some will

make no product. Short-duration batch processes minimize
the probability of mutant cell accumulation.
contamination. Certain steps involve sterile operations,
and it is challenging to maintain sterility for long periods
of time.
lot tracking. Making a product in finite batches, each
assigned a lot number, helps limit any reprocessing and/or
product recalls to a relatively small amount of product.
Portions of some bioprocesses, however, are run
pseudo-continuously. A common example is a bioprocess
utilizing mammalian cell cultures that runs in a perfusion bioreactor. This type of bioreactor is characterized by
continuous nutrient feeding and continuous cell-free broth
withdrawal. Some perfusion processes run for a month or
more. Another example is the post-fermentation centrifugation step, which separates the soluble and insoluble broth
components. This operation incorporates continuous feed of
harvested fermentation broth.
Article continues on next page
CEP

November 2009

www.aiche.org/cep

45

SBE Special Supplement:Bioprocessing

Bioprocess control considerations

Several features of commercial bioprocesses present
automation challenges:
the impracticality of directly measuring, online, what
goes on inside microscopic living cells. Rather, the control
philosophy is to (1) monitor macro externally measurable
aspects of the culture and (2) control the broth environment
in which living cells grow, replicate, and produce desired
product. For example, it is common to measure and control
dissolved oxygen and pH in the fermentation broth (i.e.,
outside the cell walls), even though the extracellular and
intracellular values of these parameters are usually different.
the uniqueness of each batch. The activity and
productivity of a batch of living cells is a function of its
history, including abnormal events that may have occurred
in upstream cell-growth operations. This batch uniqueness
causes many industrial plants to operate with a coefficient
of variability (COV) of up to 10% (or more). As a result,
changes in control setpoints should, ideally, not be made
at fixed times for each batch, but should be tied to the cell
culture achieving certain milestones, such as a target cell
mass or dissolved oxygen value.
the time-varying multi-step batch nature of the process. Start-up, shut-down, transients between steps, the
increasing magnitude of load changes as the cell culture
grows, and the management of abnormal events dominate
most of the logic in process control recipes.
responses of biological material to changes in control
setpoint or load disturbances. These responses often differ
from classical first- or second-order responses because
cells respond to changes in their environment via complex
metabolic pathways. For instance, the response to a change
in dissolved oxygen is dependent on whether the current
value of dissolved oxygen is limiting, in slight excess, in
gross excess, or close to where the cell metabolism can shift
between aerobic and anaerobic respiration.
major load changes due to the increase in cell mass.
Because the cell mass increases by several orders of
magnitude during a fermentation, the amount of agitation,
aeration, and substrate feed needed to support the cell mass
at the beginning of a fermentation can differ significantly
from that at the end of the fermentation. This phenomenon
can sometimes lead to the need for adaptive tuning and/or
split ranging of controllers.
Unfortunately, relatively few online sensors can measure
the current state of the biological culture. This is primarily
because a sensor needs to meet requirements such as:
it cannot be a potential contamination source
it must survive the sterilization step
it needs to operate for long periods of time and maintain its calibration continuously during operation
it must have high compound specificity while
46

www.aiche.org/cep

November 2009 CEP

operating in a complex broth mixture environment.

This set of circumstances a complex process and limited sensors can drive the need for alternative methods to
obtain online information (e.g., virtual sensors) and the use
of advanced control techniques.

Traditional bioreactor
measurement and basic control
Traditional online measurements for the typical bioreactor shown in Figure 2 include inlet gas flowrates, agitation
rate, tank head pressure, temperature, pH, dissolved oxygen
(dO2), and foam level. Some bioreactors include additional
sensors, such as dissolved carbon dioxide (dCO2). Most
bioreactors include some type of manual or automated
sampling system to enable certain at-line (e.g., glucose concentration) or offline (e.g., cell mass) measurements. Many
measurements are used in feedback control. A few control
loops are very simple to implement for example, a classic single-input/single-output proportional-integral (PI) loop
works well for tank head pressure control.
A more challenging example is the control of dissolved
oxygen, which is needed for most aerobic fermentations.
Measurement is typically via a single electrochemical
probe. This use of a single point source assumes a relatively homogeneous broth a questionable assumption for some fermentations, particularly those with high
viscosity. Control is normally accomplished with a cascade
configuration, with the dissolved oxygen controller serving
as the master loop providing a calculated setpoint to a slave
loop. Depending on the bioprocess, the slave loop can
involve the agitator speed, inlet air flowrate, supplemental
pure oxygen flowrate, substrate (e.g., glucose) feed rate,
and/or bioreactor head pressure. Factors such as the cells
shear sensitivity help determine the appropriate control
strategy. Agitation speed is especially common in the slave
loop for cells that are not shear-sensitive. PI control is typical for normal operation, but is often supplemented with
if-then-else rules or other techniques to override PI control in certain abnormal situations (e.g., if dissolved oxygen
becomes too low) to avoid catastrophic consequences to
the cell culture.
Another challenging control scenario occurs in many
mammalian cell culture bioprocesses (often used to manufacture proteins) in which multiple gas feeds are used to
simultaneously control dissolved oxygen, dissolved carbon
dioxide, and sometimes pH. The flowrates of the filtersterilized inlet air, pure oxygen, and pure carbon dioxide
gases are the manipulated variables, often manipulated via
an interdependent multivariable control strategy.
Each of the other bioprocess steps also offers automation
challenges and opportunities. For example, the bioreactor
sterilization step is often accomplished by raising the bio-

reactor backpressure to about 15 psig (to avoid flashing) and

controlling temperature at about 121C for 1520 min.
An automation enhancement employed in the sterilization of many bioreactors is the use of the parameter Fo,
which represents the log reductions of living organisms as a
function of temperature (T, C) and is calculated as (1):
Fo = 10[(T121)/10]

(1)

Fo increases as the sterilization phase proceeds, and

the operation terminates when Fo reaches a specified target
value. The target value is typically chosen so that the
chance of a single living (potentially contaminating) organism remaining in the bioreactor is less than one in a million.
This algorithm overcomes some of the limitations of a fixed
time-temperature strategy it includes the sterilizing that
occurs during the heat-up portion of the step, and it does
not require reaching and maintaining a specific temperature
(such as 121C, which is commonly used because the sterilization is fairly quick and that temperature is usually safe
for the equipment and broth ingredients).
Bioprocesses share an evolving paradigm with the rest
of the chemical industry the increasing availability and
use of smart sensors and valves and the use of process
control computer systems (e.g., distributed control systems
(DCSs) and programmable logic controllers (PLCs)) that
can communicate with these smart devices via digital protocols. The use of smart field devices increases the information available to automation systems, enhances online
process diagnostic capability, and facilitates the evolution
from preventive maintenance to predictive maintenance.
There is a gap between traditional bioprocess control
and optimal control. Limitations of existing bioprocesses
highlight the need for additional real-time process information. One such limitation is the use of stop-and-hold points
in processes making medicines in order to determine if the
criteria for forward processing have been satisfied. This
has significantly increased manufacturing cycle times and
increased capital requirements for equipment to hold and
store materials.
Another limitation is that the few online measurements
available provide relatively little information on the state of
the biological culture. This makes it difficult to thoroughly
understand bioprocesses and to pursue online diagnostics,
root-cause analysis, and optimal control.

A process analytical technology (PAT) perspective

Recognizing these limitations, as well as others that
apply to other portions of a pharmaceutical manufacturing
process, the U.S. Food and Drug Administration (FDA)
and industry have been working together in recent years
on a concept known as quality by design (QbD). A subset

of QbD is process analytical technology (PAT).

PAT is a continuous-improvement initiative that encourages the use of new technologies, risk-based management,
increased automation, and online data analysis to achieve a
better understanding of processes, increased online decisionmaking, reduced cycle time, improved process control, and
reduced process variation, without sacrificing quality (2).
The overall PAT vision is premised on the use of online
measurements. These range from measurement of critical process parameters such as pH and temperature with
traditional sensors, to more-complex offline assays (historically performed in a laboratory or on the production floor)
that are interfaced online or at-line to the process. Another
basis for PAT is the use of multivariate tools for online data
analysis, process control, and knowledge management,
thus further enhancing the understanding and control of the
manufacturing process.
A key objective of PAT is that the information generated
will be timely (avoiding unnecessary process delays and
holds) and can be used for appropriate real-time or nearreal-time decision-making and control. PAT is expected
to be key to the operation of a capable, compliant, robust,
in-control, and reproducible process. In other words, PAT
can be an important part of building quality into a process,
as opposed to testing for quality after the fact.
Examples of instruments that have been used to
accomplish PAT objectives include: process mass spectrometry; online high-performance liquid chromatography
(HPLC); gas chromatography; spectrometers that work in
the ultraviolet, visible, and near-infrared spectrum ranges;
and analytical systems for turbidity, pH, refractive index,
and suspended solids. PAT has already been successfully
applied to many bioprocess unit operations, including bioreactors (fermentors), centrifuges, chromatography separation columns, and tangential flow filtration.

A dynamic modeling perspective

In pursuing a better understanding of a process (a PAT
objective), a frequently used approach is to model the
dynamic behavior of a fermentations state variables in
terms of ordinary differential equations (3). The state variables for a fermentation almost always include:
cell mass
substrate concentration (e.g., the primary carbon
source for the culture)
product concentration.
A dynamic model will include a differential equation
for each of these state variables. If the broth volume varies
with time, as is typical of many fed-batch and long-duration fermentations, the model will include a state equation
for broth volume. In some cases, the model may also have
equations for the production of undesired (byproduct)
CEP November 2009

www.aiche.org/cep

47

SBE Special Supplement:Bioprocessing

d(XV)/dt = cell growth cell death

= uXV k2XV

(2)

where u = specific growth rate (h1), X = cell mass (g/L), and

V = broth volume (L). The specific growth rate, u, is a function of the theoretical maximum growth rate of a culture,
umax (h1) and the substrate concentration in the fermentation broth, S (g/L): u = (umax S)/(k1 + S). umax is unique to
each particular type of cells. The model constant k1 (g/L) is
known as the Michaelis constant and represents the value of
S at which the specific growth rate is one-half its maximum
(u = 0.5umax), and k2 (h1) is the death rate constant.
If the objective in creating the model is to determine
optimal operating conditions (e.g., substrate concentration, pH, temperature), then the model constants need to
be determined as functions of the operating parameters of
interest. In Eq. 2, only the specific growth rate dependence
on substrate concentration, u = f(S), is shown.
The state equations are dependent equations in that most
of them include terms containing one or more other state
variables. Thus, the equations cannot be solved independently, but must be solved simultaneously, usually offline.
Developing a dynamic process model invariably leads
to better understanding of the bioprocess which, in turn,
can assist in developing an optimal control strategy.
The model can be used to run trial-and-error scenarios
to determine if a higher-producing and/or lower-cost
fermentation can be achieved. This can be done in a few
minutes with the model, compared with running actual
fermentations, which can take weeks. For example, different time-varying substrate feed profiles can be explored
with the model and the best result confirmed experimentally
to determine if an optimum feed profile exists. Determining an optimal temperature profile is another optimization
opportunity, as the best temperatures for the cell growth and
product production steps are often different.
A more-advanced use of such a model is to directly
compute (rather than pursue a trial-and-error approach) the
optimal process-control parameters, such as the time-varying substrate feed rate that maximizes the amount of product at the end of the fermentation. This requires a branch of
mathematics known as calculus of variation.
Several of the bioprocess state variables, such as cell
mass (X), substrate concentration (S), and product concentration (P), are not available as online measurements for
many fermentations. If measurements or estimates of these
variables could be made online, it would improve the abil48

www.aiche.org/cep

November 2009 CEP

ity to conduct online assessment of the state of a culture,

improve the effectiveness of process diagnostics, and help
enable online process optimization.

Online models: virtual sensors

Although dynamic models based on state equations
(such as Eq. 2 for cell mass) are valuable for understanding the process and for developing control strategies, they
are typically not employed for online control of individual
batches. This is because they contain few, if any, online
measurements that characterize the uniqueness of a particular batch, and the equations are interdependent and thus take
significant computer time to solve.
Therefore, another class of models, sometimes known as
virtual (or soft) sensors, can often be utilized to predict state
variables, such as cell mass and substrate concentration.
These virtual sensors can be used in real-time and make
use of online measurements that capture the uniqueness
of a particular batch. For instance, a batch may encounter
an abnormal situation during cell growth, such as a loss of
agitator power, pH spike, or foam out, that reduces the number of viable cells. A state equation such as Eq. 2 does not
accommodate such events, but a virtual sensor can.
Implementing process mass spectrometry
on bioreactors
Process mass spectrometers are often multiplexed
to several bioreactors to measure, in near real-time, the
components of fermentation supply and exhaust gases (4).
These measurements are typically sent to a process control
computer, where the data are combined with other bioreactor measurements and information to compute parameters such as culture oxygen uptake (OU), carbon dioxide
evolution (CE), respiration quotient (RQ), and bioreactor
oxygen mass-transfer coefficient (kla). These calculated
parameters (e.g., OU and CE) can also serve as inputs to
online virtual sensor models.
Figure 3 is a sample fermentation OU trend plot. Notice
the exponential increase during the first third of the batch
Oxygen Uptake, mmol/L-min

metabolites or for additional substrates (e.g., if the culture

shifts from one primary substrate to another during the
fermentation).
A simple example of a state equation for cell mass is:

End of Exponential Growth

Time, h (or days)

S Figure 3. Oxygen uptake by the cell culture increases exponentially at

first, then follows other patterns as the fermentation progresses.

(mimicking the exponential unconstrained increase in cell

mass) and the impact of other events during the fermentation
thereafter that limit/influence the culture respiration rate.
Bioreactor applications that make use of online gas
analysis measurements include:
estimating cell mass (a state variable) and specific
growth rate
estimating and controlling substrate concentration (a
state variable)
determining the desired time (based on culture state) to
inoculate a bioreactor
providing early indication of bioreactor contamination
(for some types of cultures)
monitoring any volatile compounds produced
determining when to send exhaust gases to scrubbers
or incinerators if environmentally unfriendly gases (e.g.,
hydrogen sulfide or alcohols) are produced.
The use of process mass spectrometry to determine
culture respiration and bioreactor parameters allows for
near-continuous monitoring without perturbing the process,
as is often required for alternative methods.
Without information such as that obtained from process
mass spectrometry, there are relatively little online data that
quantitatively indicate the state of the biological culture.
Standard online measurements of aerobic bioprocess parameters, such as temperature, pH, air sparging rate, agitation
speed, and backpressure, provide very little information on
the state of the culture.
By monitoring OU, CE, and RQ, plant personnel can
continuously follow the state of the culture and know, in near
real-time, when the culture is deviating from normal historical trends. Such information is also useful for determining
when increases in physical control variables (e.g., airflow,
agitation rate, and nutrient feeds) are needed and for conducting root-cause analyses when process deviations occur.
Several virtual sensors used in industry make use of respiration data (5). The online estimation of two universally
important bioprocess state variables, cell mass and substrate
concentration, is described in the following sections.

State variable: cell mass

Cell mass is typically determined by various offline
techniques that require drawing a sample from the bioreactor
and manually performing at least some portion of the sample
preparation and/or assay. However, there is value in automatically providing a continuous online measurement or estimate of cell mass. Such an estimate can be used, for instance,
to determine when the culture in the seed tank is ready to
transfer, as the inoculum, to the bioreactor/fermentor.
For some bioprocesses, commercial sensors are available
that can be used to estimate cell mass concentration online
by measuring optical density, turbidity, NIR/IR absorption,

etc. However, these sensors may have important limitations,

including the time-varying influence of certain nutrients in
the broth (e.g., insoluble ingredients that absorb energy at
some of the same wavelengths as the cellular material), and
the inability to distinguish between viable (i.e., living) cells
that can produce product and nonviable (dead) cells.
If a practical sensor to directly measure viable cell mass
is not available for a particular bioprocess, a soft (virtual)
sensor can often be developed. Such cell mass estimators
come in many forms and degrees of complexity.
For some bioprocesses, such as slow-growing mammalian cell cultures, most respiration activity is associated with
cell maintenance, and the viable cell mass concentration (X)
is proportional to oxygen uptake and CO2 evolution:
X k3OU
X k4CE

(3a)
(3b)

where k3 and k4 are proportionality constants (g-min/mmol).

For some faster-growing cell cultures (e.g., some
E. coli and Streptomycetes fermentations), significant respiration activity is associated with cell growth. In these cases,
a model based on cell respiration activity as a function of
growth and maintenance often works well, especially during
the seed and fermentation growth phases (5). The primary
input to the model is carbon dioxide evolution, which is
assumed to result from cell growth and maintenance activities and is represented by:
CE = k5(dX/dt) + k6X

(4)

where X = mass of actively respiring (i.e., living) cells,

t = time (min), and k5 and k6 are proportionality constants
(mmol/g and mmol/g-min, respectively). If desired, OU can
be used instead of CE; the form of the equation is the same,
but the values of the constants may be different.
After fitting Eq. 4 to historical data and noting that CE
is available online, the only unknown to be computed for
the batch is the cell mass, X.
For more complex scenarios, such as those with significant respiration activity directed to activities other than
growth and maintenance (e.g., the production of product
and/or byproducts), a more-sophisticated model is usually
needed. Neural nets incorporating several measured inputs,
including OU and/or CE, often work well.

State variable: substrate concentration

The concentration of the substrate (e.g., glucose) has
historically been measured by sampling the bioreactor broth
and performing an at-line analysis or offline lab assay.
The sampling system itself can be of concern if it does not
include a sterile barrier, and it could be fouled by cells or
CEP

November 2009

www.aiche.org/cep

49

SBE Special Supplement:Bioprocessing

RG = k7OU = k8CE

(5)

where k7 and k8 are proportionality constants (and have

units of g glucose consumed/mmol OU and g glucose
consumed/mmol CE, respectively).
The glucose concentration, which is needed as the feedback signal for control purposes, can then be determined by
a standard material balance calculation incorporating RG (6).
This model is too simplistic for some cases, such as
when a fermentation generates a significant amount of
product or byproduct (e.g., acetate, lactate or alcohol) under
certain conditions, as these represent additional endpoints
for the carbon originating from glucose or other substrate.
Fortunately, there are often methods to estimate online the
magnitude of these additional carbon pathways.
For example, the amount of base added to the fermentation to control pH can be a good indicator of the metabolic
rate of acetate or lactate (i.e., in the form of acid) being
generated by the culture. For cultures producing ethanol, RQ
(respiration quotient = CE/OU) has been correlated with the
rate of alcohol production. Various models, such as neural
nets, can then estimate the substrate consumption rate based
on the available online information, including CE and/or OU.
Online models for cell mass and glucose concentration
may be accurate for short durations (e.g., hours), but will
eventually drift because of imperfect sensor calibration
and/or the use of an imperfect model. Therefore, offline
analyses of cell mass and substrate are useful for updating
online models.

Online HPLC to remove impurities from product

Lets now consider a non-bioreactor example. Online
HPLC can be very effective in the elution step of downstream chromatography product recovery and purification
operations. In this step, the desired compound elutes from
(i.e., leaves) a chromatography column at a different time
than the impurities. However, the time frame for collecting
the column output to obtain the desired compound cannot be assumed to be exactly the same from batch to batch
due to several physical and operational factors that affect
50

www.aiche.org/cep

November 2009 CEP

a columns performance (e.g., the feed rate to the column

or the use of a newly repacked column). Rather, the value
is in directly monitoring the output of the chromatography
column to determine the precise time to start collecting the
desired product and when to stop collecting it (Figure 4).
The objective is to collect as much of the desired product as possible while minimizing the inclusion of undesired
impurities. Since the chromatographic peaks of the desired
compound and impurities sometimes have overlapping tails,
the timing of starting and stopping the collection of column
output is critically important. Although direct sensors (e.g.,
UV) can be used for this purpose in some applications,
more-complex online HPLC systems are needed for others.
One pharmaceutical company has reported installing
more than 30 online HPLC systems since 1981. They are
used in applications requiring high selectivity in compound
determination and for identifying end-points (i.e., starting
and stopping) in quenching enzymatic reactions.

Advanced monitoring and control techniques

Several advanced monitoring, modeling, and control
techniques may be used on bioprocesses. These include
neural nets, principal component analysis (PCA), partial
least-squares (PLS), metabolic modeling, fuzzy logic, and
model predictive control. Reference 6 discusses several of
these techniques in detail.
Challenges and opportunities
Opportunities for improved bioprocess monitoring and
control exist in several areas (7):
improved batch process functionality in computerized
systems from automation vendors. This includes more userconfigurability (vs. hard-coding by the vendor), which is a
particular need for PLC controllers embedded in equipment
modules. In addition, automation systems should provide
Normalized Concentration

other insoluble nutrients that may be present. In addition,

there is a significant and variable delay in obtaining results
due to the time involved in transporting samples from the
bioreactor to the analyzer. This is problematic for online
control purposes for some high-respiration fermentations
(such as those utilizing E. coli bacteria), since the culture
consumes substrate at a high rate and changes in substrate
concentration can occur quickly.
As with cell mass, the relationship between glucose consumption (RG) and oxygen uptake or CO2 evolution is quite
simple, especially for bioprocess growth phases:

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2

Control recipe
to start collecting
desired product

Impurity

Main
Product

0.1
0
1 3 5 7

9 11 13 15 17 19 21 23 25 27 29
Time, min

S Figure 4. The concentration of the compounds exiting the chromatography column gives an indication of the best time to start collecting the
desired product. Source: (5). Used with permission of the International
Society of Automation.

for data and alarm record tags that include lot number and
process step and phase. It is important for these automation
systems to display information in relative time (i.e., elapsed
time since the beginning of the batch step) rather than
calendar time, because the frame of reference for monitoring batch processes and data analysis is relative time and
not time of day. Finally, the automation logic should be
consistent with the International Society of Automations
(ISAs) S-88 standard for batch processes.
more powerful utilities in historians so users can more
efficiently extract information and knowledge from process
data. The significant manual effort required to (1) decompress and review data, (2) delete outliers, (3) combine
continuous, discrete, and bill-of-material data from different
databases in different formats, (4) normalize data, (5) estimate missing data, etc., discourages data mining and the use
of such value-adding process analysis techniques as PCA
and PLS.
improved/additional online bioprocess sensors. For
example, better sensors to monitor the substrate and product
concentrations directly would be especially valuable.
improved dynamic bioprocess models. Such models
would improve process understanding and help in determining optimal time-varying control strategies.
development of hybrid neural nets. These combine
known white-box first-principles knowledge of the
process with nonlinear black-box neural-net estimators of
complex parameters (8).
publication of success stories about bioprocesses using
advanced control techniques. Engineers can learn from
the experiences of others who have successfully implemented model predictive control, fuzzy control, and other
technologies.

JOSEPH S. ALFORD retired from Eli Lilly & Co. at the end of 2006, after 35
years of employment, and is continuing his professional career with
consulting/contract work, university guest lectures, authoring contributions to the literature, reviewing technical article submissions for
several journals, and serving on several AIChE and ISA national committees. He joined Lilly to help lead the development of their fermentation
process control and data historian system, which was subsequently
replicated into all their major fermentation facilities. He subsequently
served as Lillys chief computer-systems validation auditor and as head
of the Advanced Process Technologies Group. After receiving a BS in
chemical engineering from Purdue Univ. in 1966, he served two years of
active duty in the U.S. Navy as an Engineering Ofcer on an aircraft carrier in the Vietnam War. He then earned MS and PhD degrees in chemical engineering from the Univ. of Cincinnati in 1972. He received AIChEs
Computing in Practice Award, Purdues Outstanding Chemical Engineer
Award, and several ISA awards, was inducted into Purdues Military Hall
of Fame, and was named a Distinguished Alumnus of the Univ. of Cincinnatis College of Engineering. He is a licensed Professional Engineer,
a Certied Automation Professional, and a Fellow of both AIChE and
ISA. He serves on AIChEs Publications Committee (where he chairs the
New Books Committee) and CEPs Editorial Advisory Board, as well as
several university science and engineering advisory boards.

In closing
Although much progress has occurred in recent decades
in the monitoring and control of bioprocesses, most
bioprocesses are not optimally controlled. This is due,
in part, to a lack of sufficient understanding of cellular
metabolic processes and how the external cellular environment (for which some online measurements exist) relates
to the internal cellular environment (which is not directly
measured). Consequently, most parameter control setpoints
are determined through trial-and-error experimentation.
Other factors slowing progress toward optimal control
include the limited number of sterilizable, rugged, and
stable commercially available online sensors and the high
regulatory burden (time, cost, and effort) in some pharmaceutical and biotech industries to validate new techniques
and technologies.
The increased understanding of cellular metabolic
processes (facilitated by use of well-known cell lines
such as E. coli and Chinese Hamster Ovary (CHO)),
development of improved online sensors (such as recent
improvements to dCO2 probes), improved models, and the
PAT initiative (in pursuing reduced cycle time, improved
control, and more online decision-making) should all play
important roles in the evolution toward optimally conCEP
trolled bioprocesses.

Literature Cited
1.

2.

3.
4.

5.

6.

7.

8.

Boeck, L. D., et al., Sterilization of Bioreactor Media on

the Basis of Computer-Calculated Thermal Input, Designated
as Fo, Journal of Industrial Microbiology, 3 (5), pp. 305310
(Aug. 1988).
U.S. Dept. of Health and Human Services, Food and Drug
Administration, Guidance for Industry. PAT A Framework
for Innovative Pharmaceutical Development, Manufacturing, and
Quality Assurance, www.fda.gov/downloads/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/UCM070305.pdf,
FDA, Rockville, MD (Sept. 2004).
Leigh, J., Modelling and Control of Fermentation Processes,
P. Peregrinus Ltd., London, U.K. (1987).
Alford, J. S., Bioprocess Mass Spectrometry: A PAT Application, Journal of Process Analytical Technology, 3 (3), pp. 610
(May-June 2006).
Buckbee G., and J. Alford, Automation Applications in BioPharmaceuticals, International Society of Automation (ISA),
Research Triangle Park, NC (2008).
Boudreau, M., and G. McMillan, New Directions in Bioprocess Modeling and Control, International Society of Automation
(ISA), Research Triangle Park, NC (2007).
Alford, J. S., Bioprocess Control: Advances and Challenges,
Computers and Chemical Engineering, 30 (1112),
pp. 14641475 (Sept. 2006).
Psichogios, D., and L. Ungar, A Hybrid Neural Network
First Principles Approach to Process Modeling, AIChE Journal,
38 (10), pp. 14991511 (Oct. 1992).

CEP

November 2009

www.aiche.org/cep

51