Robust optimization of cascaded MSMPR crystallization unit using unsupervised machine learning

Abstract

AbstractThe use of mixed suspension mixed product removal (MSMPR) system in the pharmaceutical industry to produce active pharmaceutical ingredients is well known. In industrial settings, the MSMPR system is subject to lot of process uncertainty which, if ignored, might result in poor product quality. In this work, the process uncertainty involved in MSMPR is targeted during the process optimization stage to find robust optimal operating conditions. The temperature and the residence time inside each cascaded MSMPR unit, altogether six, are considered as uncertain parameters. A sampled set of uncertain data points for such six different uncertain parameters are clustered using a novel support vector clustering (SVC) based algorithm. The uniqueness of this algorithm lies in its ability to fine‐tune the hyper‐parameters of SVC while intelligently clustering the uncertain data points into optimal number of clusters. Such identified clusters are helpful to generate more samples from the intended regions rather than generating them randomly to avoid proposing conservative solutions. Both best‐case and worst‐case scenarios for robust oOptimization (RO) are considered with , and samples. As the model has to be evaluated for a large number of samples and the MSMPR models are time‐consuming to evaluate, a surrogate model of the MSMPR process is developed to perform optimization under uncertainty. Performance metrics are used to quantitatively establish the superiority of the SVC based RO over the box‐sampling based RO.