Moving from individual continuous unit operations to fully connected biopharmaceutical manufacturing will require greater coordination between sensors, models, and control systems, according to a new study.
The review argues that continuous biomanufacturing is fundamentally a process-control and system-integration challenge. Unlike batch production, continuous processes must remain within defined operating and quality limits over extended periods while responding to equipment drift, disturbances, and interactions between connected upstream and downstream operations.
Conventional proportional-integral-derivative control remains well suited to stable, clearly measured variables such as pH, pressure, and flow rate. However, it is less effective when several process variables interact or when measurements are delayed. Model predictive control can address these more complex situations by forecasting process behavior and adjusting multiple inputs while respecting operating constraints.
The review describes applications spanning perfusion cell culture, chromatography, viral inactivation, viral filtration, and single-pass tangential flow filtration. In one cited six-day filtration study, model-guided feedback control maintained monoclonal antibody concentrations within 0.5–2.5 percent of their setpoints and kept buffer exchange within 8 percent of its target.
Process analytical technology provides the measurements needed to support such control. Although in-line and online sensors offer frequent information, the authors note that they can be costly to install and maintain and may be affected by calibration drift. At-line or automated sampling systems may provide sufficient measurement frequency for slower processes while offering greater analytical accuracy.
When critical attributes cannot be measured directly, soft sensors can estimate them from other process data. Their predictions remain vulnerable to sensor noise and model error, however, creating a risk that inaccurate estimates could trigger inappropriate control actions.
Reinforcement learning could eventually support adaptive process control but remains largely confined to simulations and laboratory demonstrations. Bioprocess experiments are expensive and slow, limiting the data available for training. The need to explore different actions during learning also presents safety and regulatory problems, while limited interpretability makes it difficult to explain why a controller selected a particular action.
Digital twins could provide a lower-risk environment for testing these strategies. By reproducing process behavior virtually, they could help manufacturers evaluate disturbances, optimize operating conditions, and train control algorithms without endangering a physical production run.
The authors propose a gradual route toward implementation. Initially, digital twins could detect deviations and support operator decisions within predefined limits. With sufficient validation, they might progress to supervisory optimization and eventually make controlled adjustments themselves. However, more sophisticated control is not always the best choice. Advanced systems require additional sensors, connectivity, model-development experiments, validation, maintenance, and change control. Their use must therefore be justified by measurable improvements in product recovery, consistency, or risk management rather than technological capability alone.
Newsletters
Receive the latest pharmaceutical news, personalities, education, and career development – weekly to your inbox.
