Applying an MPC algorithm to the dissolved oxygen level of an intermittent fed-batch process

Providing cells with ideal physiological conditions for a respective goal is one of the main tasks of a bioreactor. The level of dissolved oxygen (DO) in the broth plays a pivotal role here, as limitations cause shifts in the metabolic activity of the cultivated organisms. These can lead to byproduct accumulation or even cell death. Therefore, controllers are employed to stay above a certain threshold by acting on the agitation, air flow and oxygen partial pressure, where many are based on PID algorithms. In order to counter the nonlinear nature of bioprocesses they were extended by implementing feedback linearization, cascaded controllers or gain scheduling (Babuška et al. 2003). Nevertheless, the downside of their reactive principle can be seen when the system is faced with abrupt changes in nutrient addition. Such conditions can be observed during the transition of phases in a process or because of intermittent feeding profiles in high-throughput small scale multi-reactor systems. The shot-wise addition of the substrate results in sudden drops of the DO signal as described by Kim et al. (2023) and eventually surpass the threshold of the system due to the delayed response of the control loop. The following oxygen limitation can have considerable effects on the health and productivity of the organism, creating a need for better control.

This work explores the potential of a predictive model-based algorithm to prevent oxygen limitations in an intermittent fed-batch process. The oxygen uptake is modeled through the metabolic activity of an aerobic microbial process by employing simple growth kinetics and elemental balances. This demand is met with the oxygen transfer rate, which is connected through a kLA model established by Van’t Riet (1979) to the two actuators of the control loop, the agitation speed and the aeration rate. The parameters of the resulting models are estimated in a laboratory bioreactor setup, using a combination of online and offline signals, as well as the concentrations of the off-gas. The model is then implemented and assessed first as a soft-sensor during a process that simulates an intermittent feeding profile, before incorporating it into an MPC algorithm. The performance of this predictive control strategy is then evaluated and compared to other established technologies, such as PID control. The predictive nature of the MPC is expected to prevent sudden drops in dissolved oxygen and thereby presents a promising control algorithm for small scale reactor systems with intermittent feeding or processes with harsh phase transitions. Since the model is based on simple growth kinetics, it is generically applicable to aerobic cultivations at different scales after parameterizing the reactor specific kLA model.
Acknowledgement