Continuous pharmaceutical manufacturing (CM) demands advanced control strategies capable of operating reliably under real-world measurement noise, process variability, and ill-conditioned experimental data, conditions where classical system identification methods frequently fail to generalize. In this work, regularized Dynamic Mode Decomposition with Control (rDMDc) is proposed as a robust framework for data-driven system identification and predictive control, incorporating Tikhonov regularization to stabilize model identification, suppress noise-induced overfitting and provide reliable linear state-space models suitable for real-time predictive control. The framework is demonstrated on two case studies: a continuous production of 2-nitrodiphenylamine (Li-NDPA), where rDMDc outperforms DMDc, noise-handling variants and N4SID methods with a RMSE of 0.026 and a lower uncertainty 95% predictive interval length of only 0.18 and a segmented fluidized bed drying operation in a ConsiGma™-25 Continuous Tableting Line reducing validation RMSE by 26% for the loading phase (2.35 vs. 3.19) and by 63% relative to DMDc for the drying phase (0.85 vs. 2.34). rDMDc models enable accurate dryer temperature and moisture control, demonstrating superior setpoint tracking and disturbance rejection in an MPC framework. Our results demonstrate that rDMDc provides stable, interpretable, and generalizable models for continuous pharmaceutical processes, enabling efficient model-based control under real-world noise measurement.
Vega-Zambrano et al. (Mon,) studied this question.