An industry-standard approach to ensuring the robustness of biological drug substance involves replicating plant-scale temperature exposure in the laboratory. Achieving this requires an accurate dynamic model of the chiller/heater system in a digital environment. In this work, a System Identification approach is demonstrated to accurately map the coolant setpoint (input) to the product temperature (output). Input-output data are collected for a dynamic chiller system with PID feedback control, and the performance of various models is compared including transfer-function based process models, nonlinear autoregressive exogeneous models, and dynamic neural networks. To estimate the unknown input in the physical domain, two configurations are tested in the digital domain: a PID feedback loop and inverse modeling. Training and validation results indicate that dynamic neural networks effectively capture the system dynamics for unknown input identification. A key limitation of this approach is the need to pre-process the input data to remove noise, particularly exothermic freezing peaks. Overall, this work demonstrates a practical, data-based modeling approach to ensure the robustness of biological drug substance.
