Risk Assessment for a Twin-Screw Granulation Process Using a Supervised Physics-Constrained Auto-encoder and Support Vector Machine Framework.

Quality risk management is an important task when it pertains to the pharmaceutical industry, as this is directly related to product performance. With the ICH Q9 guidelines, several regulatory bodies have encouraged the pharmaceutical industry to implement risk management plans using scientific and systemic approaches such as quality-by-design to asses product quality. However, the implementation of such methods has been challenging as assessment of risks requires accurate quantitative models to predict changes in quality when variations occur. This study describes a framework that quantitatively assesses risk for a twin screw wet granulation process. This framework consists of a physics-constrained autoencoder system, whose outputs are constrained using physics-based boundary conditions. The latent variables obtained from the auto-encoder are used in a support vector machine-based classifier to understand the granule growth behavior occurring within the system. This framework is able to predict the process outcomes with 86% accuracy and classify the granule growth regimes with a true positive rate of 0.73. Based on the classification the risk associated with the process can be estimated.