Reactive distillation processes used in monomer production for high-performance adhesives require stringent control of critical quality attributes (CQAs), yet they are highly sensitive to feed variability, uncertain kinetics, and nonlinear reaction separation interactions. Because CQAs such as purity and viscosity are traditionally measured only through offline laboratory analyses, quality deviations are detected too late for effective intervention, leading to extended cycle times, off-spec material, and elevated resource consumption. This motivates the need for real-time, model-based approaches capable of predicting product quality throughout the batch.
This work presents a hybrid digital twin that integrates mechanistic reactive-distillation models with machine learning to capture complex behaviours not fully described by first principles. Historical process data, laboratory measurements, and equipment specifications were combined to develop soft-sensing models that estimate key CQAs continuously. Validation results demonstrated over 94% accuracy for both temperature trajectories and purity predictions across varying operating conditions. Deployment of the digital twin enabled real-time quality prediction, improving endpoint determination and reducing cycle time by 29%. A deviation-detection framework supported proactive identification of off-normal trends, contributing to an 80% reduction in batch failures.