Header Image - Design to improve.
10 Articles

What is a Soft Sensor or Software Sensor?

Soft Sensor Highlights

  • What is a soft sensor (software sensor or softsensor) and how does it differentiate from a hardware sensor?
  • Where can you use soft sensors in industrial bioprocessing?
  • How can you measure the biomass concentration in bioprocesses using soft sensors?
  • How do you get started with soft sensors in your bioprocessing lab or manufacturing facility?
  • How to run soft sensors in inCyght bioprocess analysis software

Soft Sensor Benefits

  • Soft sensors make the most out of your data and signals that you have already collected in your bioprocess
  • Perform real-time analyzing, monitoring and control
  • Reliant calculation of parameters where no hardware sensor is available
  • Reduce purchase and maintenance costs
  • Reduce risk of contamination of the bioreactor

You are a bioprocess professional, you frequently hear and read about “soft sensors” and you get the idea you should already know about them? You keep wondering what people are doing with soft sensors in their locked-up laboratories and manufacturing facilities? Congratulations, this post is for you!

Soft Sensor Definition

The Measurement, Monitoring, Modelling and Control (M3C), a working group of ESBES, explains the origin of the name softsensors:

“The term combines the words “software”, because the sensor signal evaluation models are usually implemented in computer programs, and “sensors”, because these models are delivering information similar to hardware sensors”

The key here is “similar to hardware sensors”. It’s what changes this definition from one that could describe almost any form of calculation performed on sensor data. A soft sensor creates a new, hardware-sensor like signal. The signal is produced by a software, instead of a hardware sensor. Like every other measurement signal, the soft sensor signal is used for analyzing, monitoring and/ or controlling your bioprocesses.

You can distinguish between a softsensor and an ordinary calculation on sensor data, if you ask yourself the question “Is this a new hardware sensor-like measurement or just a refinement of a signal I already use?”

At this point, despite this definition and explanation, you’re probably still wondering what exactly a soft sensor is. In microbial bioprocesses, you measure the biomass concentration to initiate process events (e.g. induction) and, following best practices for the analysis of bioprocess data, to identify processing trends. You can measure the biomass concentration by sampling your bioreactor, centrifuging your sample and gravimetrically measuring the biomass dry cell weight. You can also estimate the biomass dry cell weight concentration by in-line sensors based on back-scattering or dielectric spectroscopy (permittivity). But what can you do if your budget does not cover the purchase?

A soft sensor to measure the biomass concentration

You can measure the biomass concentration with a soft sensor based on off-gas measurement. The big advantage is that this is non-invasive and does not require an additional port to your bioreactor. To predict the biomass concentrations from the CO2 and O2 measurements of your off-gas analyzer as e.g. BlueInOne, you need a microbial soft sensor. For the calculation of the biomass concentration, the microbial soft sensor uses data from off-gas analysis, mass flow controllers and feed rates. Based on parameters such as biomass composition, taken from a host stoichiometry library, the soft-sensor calculates the biomass concentrations, specific rates and yield coefficients as outputs. The applied algorithm for the estimation is based on an equation system using multiple balances, for details please see Wechselberger et al.

Soft sensor provides a biomass signal in real-time and turnover rates

The output of a microbial softsensor is the biomass concentration, which is predicted in real-time in an interval of 2 minutes. Next to the biomass dry cell weight the microbial softsensor provides volumetric turnover rates (oxygen uptake rates, carbon dioxide evolution rates, substrate uptake rates, biomass formation rates) and corresponding specific rates (specific growth rate, specific substrate uptake rate, etc.). You can use the turnover rates to monitor the progress of your process in real time and initiate process events. Of course, like normal hardware sensor measurements, you can use all signals for control loops, for example to control the specific growth rate or a golden batch trajectory.

To run soft sensors on historical data for analytics, you require a state of the art bioprocess analytics software such as Exputec inCyght. The biomass soft sensor is also featured in the bioprocess control software BlueVis from BlueSens.

Soft sensor for many organisms

Microbial softsensor from inCyht® are applicable for processes in most common production hosts such as E. coli or P. Pastoris and will be continuously extended.

Soft sensor mode of operation

You can use the microbial softsensor in a range between 8 and 80 g/l biomass dry cell weight which is equal to 32 to 320 g/l biomass wet cell weight. This covers the full biomass range of microbial fed-batch processes. The error you can expect here is about 5 to 10% without prior calibration. You can reduce the error by performing a calibration to your specific process.

Soft sensor calibration

There is no obligation to calibrate the software sensor. However, you can improve the accuracy by performing a calibration to OD or dry cell weight measurements.

Soft sensor benefits

Perhaps more important than to understand how the soft sensors algorithms work, is to understand why soft sensors are essential to your laboratory or manufacturing facility. Traditional hard-type sensors are great when it comes to generating high-quality data. Today, state of the art bioreactor instrumentation are pH, dissolved oxygen, temperature, quantification of feeds and off-gas analysis. This progress in robust sensor technology in the last years was the basis on the global success of soft sensors.

Softsensor summary

To put it simply, soft sensors make the most out of your data and signals that you have already collected in your bioprocess. Instead of considering to buy more and more hardware, you should decide to use as much as possible soft sensors connected to your existing sensors your bioreactor. Using soft sensors you will benefit by:

  1. Perform real-time analyzing, monitoring and control
  2. Reliant calculation of parameters where no hardware sensor is available
  3. Reduce purchase and maintenance costs
  4. Reduce risk of contamination of the bioreactor

Take action in the moment by using inCyght® to analyzing, monitoring and control your bioprocess process with the next generation of smart softsensors.

Test the soft sensor in ten minutes

Want to test the microbial soft sensor on your own process data in ten minutes? As BlueSens and Exputec customers have the exclusive opportunity to use the bluesens.incyght.com webservice to test yourself how good the microbial soft sensor works on your E. coli or P. Pastoris fed-batch process data. You can contact contact@exputec.com to test it yourself.


Best Practices for the Analysis of Bioprocess Fermentation Data


  • Approved methods for fermentation data management & analysis
  • State of the art technologies for the identification of processing trends
  • Best practices for statistical analysis of performed scale-ups and technology transfers
  • Use of soft sensors in bioprocess analysis workflows


  • Traceable detection of processing trends
  • Avoid inefficient repeats and iterative process development cycles
  • Compliant in the scientific analysis of bioprocess data

As a fermentation scientist or process development manager, you are responsible for the efficient development of high performing bioprocesses, as well as robust and predictable scale-up for future manufacturing in large scale. Consistent data management and data analysis plays a critical role for the success of your development and manufacturing goals. If you  are working in the field of bioprocess development, scale-up process validation or manufacturing excellence, you might frequently ask yourself: How do I get a complete and reliable overview of the process data with less effort? How to structure the ample amounts of data from different sensors in a single database? Which methods shall I use for data analytics and the statistical evaluation of bioprocesses? In this article we will give a brief overview on bioprocess data management, bioprocess data visualization, and statistics to evaluate bioprocesses.

Data Management

If you are involved with fermentation processes, you deal with large amounts of sensor data (e.g. pH, temperature and dissolved oxygen measurements), product quality data (e.g. product concentrations, specific activity, relative potencies), “non-numerical” data such as pictures (e.g. scanned SDS-PAGE data) and many more. For every analysis purpose, you need to reorganize the data in a time consuming process manually from different data sources. However the expectation is, that a fermentation scientist should search through batches quickly to screen for fermenter-type, media name, product or project type. Best practice database requirements are i) a suitable database model to store all bioprocess relevant data in one common database, ii) the possibility to assign phase information to time-series data and iii) database filters to identify batches you want to analyze quickly.Commonly used database filters in bioprocesses are:

  • Media Type
  • Process Type
  • Strain
  • Project
  • Customer
  • Product
  • Development stage
  • Start Date/ End Date
  • Operator/ User

Visualize Data

“You can see a lot just by looking”  – Yogi Berra. Yogi Berra and more recently Nathan McNight from Genentech nailed it: For bioprocess scientists, visualization is the most important tool to detect processing trends. See his great presentation at the CMC Strategy Forum which is available online here. 
To detect trends fast and report the results, you want to create your scientific visualizations directly from a common database, without the need for exporting the data or manually manipulating data in spreadsheets – which is prone to handling errors and not recommended. Visualizations, like multi-axis overlay plots to analyze processing trends, and boxplots/ histograms to relate quality and product attributes, are excellent for viewing your data. Commonly used visualization tools for bioprocesses are:

  • Multi-axis overlay plots
  • Bar-Graphs
  • Histograms

Statistical Data Analysis

Creating nice visualizations is a good way to draw conclusions from fermentation data. However, when communicating to management or to regulatory authorities there is the requirement to statistically verify the visualized trends. The communication of process development results to regulatory authorities (process validation stage 1) has recently become very statistics-driven . This is the point where you want to have realiable and simple, but also powerful statistics, based on a common database in Exputec’s inCyght. Below you can find a brief excerpt of statistical tools that are frequently used within the bioprocess lifecycle. See also an example below on how statistical equivalence testing can be realized.

Commonly used statistical tools

  • Statistical equivalence testing: Is the large scale process similar to my small scale process?
  • Statistical Power analysis: How many experiments do I need for my experimental study?
  • Statistical tests/ regression modelling: Do I see a significant impact of one factor on my fermentation performance?

Soft Sensors for Information Mining

Soft sensors make the most out of your data and signals that you have already collected in your bioprocess. You can run soft sensors in real-time or on historical data using a bioprocess analytics software such as Exputec inCyght. Instead of considering to buy more and more hardware, you should decide to use as much as possible soft sensors connected to your existing sensors your bioreactor. Learn more about the use of soft sensors here.



How to realize data management, visualizations and must-have statistics, is one of the most important question a fermentation scientist or process development manager has to answer. One possibility is to use spreadsheets. Although this might sound like a quick fix, you will run into serious difficulties on the short and long run. You will miss the possibility for intelligent ways to search for batches and filter for the data you really need. The calculations are time consuming  and  rarely reproducible by another scientist or for management review. You will spend a lot of time trying to standardize your Excel sheets using formulas and macros, and at the end it is very likely that they are inconsistent. On the long run, you will lose a lot of time copying and pasting data from spreadsheet to spreadsheet.

A second possibility is to set-up a fermentation database for your company and connect it with statistical and visualization software. Various data sources with different data models and data formats have to be connected the database. The data model will be very complex, as you have to ensure to store additional information such as  process phases, etc… You must maintain the software and you will have a costly patchwork of many different software components without the possibility of direct database integration.

The best practice is to rely on an out-of-the box solution for bioprocess data management, visualization and statistic. Increase operational efficiency and transparency with inCyght®, a standardized software solution for the leading bioprocess companies around the world. Digitize, and automate your data management, visualization and try our feature on scale-up analysis for fermentation processes. Improve your compliance and join the best practice of leading bioprocess companies:, contact us at contact@exputec.com to receive more information and a personal demo.

Strategies for mechanistic modelling of bioprocesses

In a discussion on LinkeIn we were asked how Exputec approaches mechanistic model building in a fast and efficient way. Here our reply:

How to approach model building of course depends on the goal of the modelling and the project resources (time!) available in the project. Time for the model building time is always is a critical factor, therefore we need to be fast. To achieve this, we have a system of i) a model library for standard models for microbial/ cell culture fermentation, primary recovery and chromatographic process models in ODE form (sets of differential equations). Based on this model library, observabilities are checked and kinetics are trained based on reference data. So we end up with hybdrid models, where the general states are described in ODEs and complex kinetics are modelled using a data driven approach.

This enables us to realize e.g. microbial feeding profile optimizations or to implement cell culture control algorithms (glucose/ glutamine control) in a few weeks of overall project time.

Predictive analytical tools for bioprocess design-, analysis and control

Following an intensive discussion on LinkedIn about predictive analytics for the bioprocess industry, here a comment on strategies and tools that are used by the biotech industry:

The predictive analytical tools used for bioprocess design, analysis and control need to be differentiated:

First, there are information mining tools that are used in combination with multivariate and statistical data analysis. Relevant data from bioprocesses is mined, e.g. phase durations, growth characteristics, material attributes etc.) and evaluated (statistical workflows, MVDA tools) for process optimization, process trending, similarity analysis etc.

This is state of the art in the biotech industry and also a regulatory requirement for process validation as regards biopharma processes, see the 2011 process validation guideline. The Exputec inCyght bioprocess analysis software has integrated the full functionality necessary for this tasks (data management, information mining, multivariate/ statistical analysis/ reporting) and is used extensively for this purpose by biotech customers.

Second, there is mechanistic modelling, e.g. ODE models, applications in industry are e.g. microbial feeding profile optimizations, cell culture control algorithms for the control of glucose/ glutamine (especially relevant for biosimilars) or mechanistic soft-sensors.



Statistical- and Mechanistic Experimental Design

by Patrick Sagmeister 1 Comment

The classical “Statistical DoE” applies experimental screening- and optimization designs templates for the screening, optimization and robustness testing of processes. This type of DoE is used in case no or little prior process knowledge is available. However, since for most industrial processes a high degree of knowledge is available, statistical DoE is typically not the most efficient strategy. Technologies are needed to efficiently incorporate the available knowledge on the process to increase the experimentation efficiency. Mechanistic DoE, or M-DoE, incorporates engineering knowledge in the form of constraints, expert-rules and known mechanistic relationships to efficiently experiments to identify optimal process conditions faster. Thereby, process optima are identified using less experiments and DoE design pitfalls are avoided.


Tasks of Manufacturing Data Scientists

by Thomas Zahel 0 Comments

Manufacturing data science is the discipline of extracting knowledge from large volumes of manufacturing data. As data scientists, we solve challenges such as the reduction of failed batches, root cause analysis for unexplained manufacturing deviations and the optimization of manufacturing processes.

A typical project starts with a detailed discussion of the manufacturing process along process flow, block-flow diagrams and risk assessment documents. In customer workshops, we specify the goals of the analysis and discuss open issues and the available data sources. Frequently, we also visit the manufacturing process on-site. Thereby, we gain a deep understanding of the manufacturing process which is necessary for the efficient analysis using data science tools.

Next steps are the transfer of the manufacturing data to be analyzed. We retrieve data from databases, MES, LIMS systems or multiple Excel spreadsheets. Sometimes, our client provides us with unstructured data in paper form, for example batch records. Then, we digitalizes data from batch records available in paper form. Anyway, we find a way to get a grab on the data. As soon as we have a grab on the data, we apply our data science approach based on the inCyght data science framework: Information is mined from large datasets using our algorithm database, including software sensors, and analyzed using multivariate methods. This enables to develop clear and interpretable hypothesis for root causes. inCyght information mining makes it possible to display information – already present in huge amounts of data – in a clear and meaningful way in order to make interpretations possible.

Following this framework we identify the targeted root causes. This is the most rewarding step, when we finally get an understanding of the variances in the data and can trace a manufacturing challenge back to process variables. Once we have identified the underlying root causes, we set up customer workshops to discuss preventive- or optimization strategies to overcome the manufacturing challenges.

Summarizing, you see that a manufacturing data science is not a job that is performed closed doors. We interact with our customers and the manufacturing processes on a daily basis. This requires, next to data science skills,  inter-personal skills and engineering know-how.


What is a Software Sensor?

by Patrick Sagmeister 1 Comment

A software sensor, or short “soft-sensor”, is a software that is capable of predicting non-measured process variables based on a mathematical model. A typical application of a software sensor is the online-monitoring of the biomass concentration and the specific growth rate in bioprocesses. Benefits of using software sensors compared to traditional “hardtype” sensors in the field of bioprocessing are minimal investment and maintenance costs and no violation of the sterile barrier of the bioreactor. Furthermore, software sensors can also be used to measure their entities in historical datasets.

Today, software sensors are frequently integrated in process control loops, used for routine process monitoring and as information mining tool in course of process optimization.

How to align Manufacturing Data?

As manufacturing data scientists we are facing piles of unstructured data form different data sources. If we are lucky, the data to complete a data science project is located in a nice manufacturing database. However, more commonly, we have to tap user-created Excel documents, pdf records or partly structured data retrieved from MES- or LIMS systems. This is partly due to the fact that we are tapping data created decades ago. So, in case you intend to go into manufacturing data science, don’t be suprised to to retrieve data from a magnet storage from time to time. So, for the appripriate alignment of manufacturing data it is crucial to unserstand pitfalls of data import and alignment from different data sources.

What is Manufacturing Data Science?

Manufacturing data science refers to the purposeful application of data science and machine learning methods to solve manufacturing challanges.

Today, we are witnessing the fourth industrial revolution (Industry 4.0), which is accopanied by the massive collection and analysis of manufacturing data. Using an appropriate data science framework, companies act on manufacturing data insighty to improve productivity and product quality.

Typical use cases for manufacturing data science are

  1. process optimization,
  2. process troubleshooting and
  3. the support of operational- and process design decisions based on manufacturing data insights.

Software Sensors for Process Monitoring

by Valentin Steinwandter 0 Comments

The knowledge extraction out of bioprocesses requires the measurement of multiple process parameters. Those parameters can be classified depending on how they are measured; on-line, at-line or off-line. On-line variables are measured by sensors directly in or on the reactor, e.g. pH, dissolved oxygen, temperature or offgas composition. At-line measurements are automatic measurements directly at the process, e.g. different spectroscopic methods like infrared or fluorescence spectroscopy. Off-line measurements are executed manually. Examples therefore are the measurement of dry cell weight (biomass), or the analysis of chemical components through HPLC.

Off-line variables carry important information about the process. However, often they are complicated, error-prone and – as one of the largest drawbacks – time delayed. Because of those drawbacks the interest in software sensors (short: softsensors) increased during the last decades. Softsensors are a combination of physical sensors (e.g. offgas analyser) and algorithms (chemometric models). Jointly they can be used to estimate off-line parameters like chemical compounds or biomass in real-time. They do not require additional physical instruments but combine the already existing process data to new variables.

Softsensors require an excellent on-line measurement equipment as well as exact calibration thereof. Furthermore, a good knowledge about the used organisms and substrates is useful as some assumptions and/or calibrations are required when configuring the softsensor.