(164b) Digital Twins for Continuous Manufacturing in Pharma: The Case of Dissolved Oxygen Control

The Continuous Manufacturing (CM) process provides a self-contained, single-use, reconfigurable platform, that incorporates principles of automated and continuous bioprocessing. For maintaining a steady state of biomass, effective process monitoring, and controls must be in place to ensure accurate cell density, critical nutrients and metabolites, as well as the relationship between growth and metabolic data. Measurable quantities that are typically leveraged are pH, temperature, and dissolved oxygen (DO) concentration.

An original digital twin platform is proposed for plug-and-play interaction with the existing closed loop regulation of DO in the bioreactor. It performs critical parameters’ identification and enables synthesis of advanced control strategies for improved regulation of DO concentration levels. The platform is designed to cope with real-world challenges at the manufacturing level, that include but are not limited to corrupted data, unmeasurable disturbances, insufficient knowledge of process dynamics, modelling gaps and limited space for experimental designs. The platform consists of a blend of computation and processing components, that perform sequential calculations for on-the-fly, rapid learning, and system identification. It performs joint system identification and advanced control with respect to regulation of process of interest.

Our case study considers a fictitious problem of regulation of DO concentration levels in bioprocess setup that mimics Merck’s CM production line. We demonstrate how the proposed digital twin architecture combines heterogeneous data from the established closed-loop process, informs the mass balance model that dynamic DO concentration follows. The identification part provides estimates of constant (e.g. kLa coefficients, time-delays etc.) as well as non-constant parameters, such as specific oxygen uptake rate. System identification is essentially an intermediate step of an inverse engineering task that delivers an informed digital twin model. The latter model will then be the basis to design model reference advanced controls. We show that the proposed Digitally Twin-enabled joint dynamic controller outperforms the existing PID control structures.