Algorithms that respond to real-time feedback and can learn to respond to experimental irregularities can help deal with the complexities of biological manufacturing. That’s the subject of a talk due to be given at Bioprocessing Summit Europe in March 2024.
The presentation, by Nadav Bar, PhD, professor of chemical engineering at the Norwegian University of Science and Technology, focuses on Model Predictive Control (MPC)—the use of algorithms to optimize a process, rather than just correct for process deviation.
“We’re using a model predictive control algorithm to balance microbial bioprocesses to minimize resources while maximizing the product,” he explains.
According to Bar, MPC is used in many non-biological devices, such as mobile telephones, to perform processes in a user-defined optimal way. Hidden inside the MPC algorithm is a mathematical model that guides control decisions. In contrast, conventional bioprocessing might control only basic parameters, such as the amount of oxygen introduced into the bioreactor.
However, as he explains, this is often not enough because too many resources may be spent growing cells. Alternatively, there may be too few cells to produce the optimal amount of product.
In contrast, the group’s MPC strategy uses measurements of chemical compounds, such as the sugar content of the media, to work out whether cells are likely to starve of low sugar or, alternatively, die from over-high sugar content.
Chemistry measurements
“It’s a delicate balance and you need to take measurements of the chemistry inside the bioreactor to maximize [production of] the product,” according to Bar. “Such measurements are not new but are often not as rapid as required by a typical MPC controller.”
The model behind the MPC algorithm is kept as simple as possible because “it is always wrong” due to the complex chemistry inside the bioreactor. To keep the model on track, it’s necessary to provide rapid feedback on the error between model predictions about bioreactor chemistry and the reality.
“It’s a repetitive process,” he says. “And it’s new because we have the computational power to do it quickly and with technology to receive real-time measurements.”
The team is using a variety of real-time measurements from advanced instruments to provide feedback to the process control model. Where the measurements are less frequent than desirable, he explains, they have algorithms to compensate.
They are also working on adjusting the model to deal with complex sugars, not just glucose, and to deal with unreliability and noise in their data. Among other innovations, they’ve connected an MPC model to a machine learning algorithm that updates the model to deal with unexpected data.
“Traditionally we used to discard experiments where there were irregularities,” he points out. “But [over the last couple of years] we’ve been sharing and learning about the irregularities, so our model knows how to behave. Our models, and therefore the MPC, then learn with the help of the machine learning algorithms.”
