Model Predictive Control to Avoid Oxygen Limitations in Microbial Cultivations - A Comparative Simulation Study

In cell cultivation, the physiological conditions inside the reactor are critical for achieving the best overall process performance. To reach high titers and productivity, bioprocess engineers try to provide the microorganisms with the best possible environment for the given objective. The dissolved oxygen (DO) level is integral to this, as limitations cause shifts in the metabolic activity of the cultivated organisms or even cell death. This process parameter is commonly manipulated by the stirring speed and the aeration flow rate, where controllers are employed to keep it above a certain threshold. Often simple PID algorithms are deployed for this task, which are then extended with feedback linearization, cascaded control or gain scheduling to tackle the inherent nonlinear nature of bioprocesses (Babuška et al. 2003). Still, when faced with abrupt changes in nutrient addition, the purely reactive nature these systems results in sudden drops of the DO signal and extended periods of oxygen limitation. This problem is encountered in systems where the substrate is added with intermittent bolus shots, such as high-throughput small scale multi-reactor systems paired with a pipetting robot, as described by Kim et al. (2023). These metabolic challenges for the cells can affect their physiological health and further the productivity of the process, which gives the need for a more advanced control scheme.

Model Predictive Control (MPC) emerges as a promising alternative to prevent oxygen limitations using a rather simple model for the non-linear process dynamics and providing it with the known feed trajectory. The resulting oxygen uptake is modeled by combining first principal mass balances and a Monod-type kinetic for the metabolic activity. The oxygen uptake rate is then connected to the physical oxygen transfer rate of the provided air into the liquid phase through the kLA correlation proposed by Van’t Riet (1979). The parameter estimation is done with lab-scale experiments, where a combination of offline and online measurements is recorded. The MPC algorithm is then tested in-silico with different configurations for the objective function and compared to the performance of a common PID control. The simulated performance of the predictive controller shows that the time of oxygen limitation throughout the process can be minimized by anticipating the needed control action before the next feed pulse is added. These results show that the proposed algorithm can ensure an improved dissolved oxygen level throughout the changing dynamics of the bioprocess, even when challenged with sudden changes in nutrient supply