Development of a Fast-Solving Machine Learning Surrogate Model for a Pharmaceutical Manufacturing Digital Twin
High-fidelity models (HFMs) for twin-screw wet granulation (TSWG) are often too computationally expensiv//e for routine calibration, optimization, and digital twin deployment. This chapter presents a faster, cheaper, and easier-to-use surrogate modeling workflow that preserves HFM-level predictive capability. Sensitivity analysis is applied to prioritize influential inputs, and a rate mechanism global system analysis reduces the calibration space by 60%, leaving only 10 critical parameters. Using this reduced set, a compact mechanistic surrogate comprising 10 equations achieves accurate particle size distribution (PSD) predictions with an R of 0.998 and a simulation time of 0.41 seconds. In addition, a hybrid TSWG model is developed by coupling a population balance model (PBM) with an artificial neural network (ANN). The ANN provides rapid first-pass estimates of kinetic rate parameters within the calibration range, which are then integrated into the PBM to refine PSD predictions and accelerate calibration. The proposed workflow, adaptive framework for robust and accelerated model evaluation (AFRAME), offers a structured and efficient methodology for model calibration in continuous pharmaceutical manufacturing and supports the development of robust, adaptive digital twins applicable to a broader range of manufacturing processes.