In Situ Imaging Combined with Deep Learning for Crystallization Process Monitoring: Application to Cephalexin Production

Abstract

The online detection of a trace amount of an undesired solid phase within a crystal slurry can enable feedback control to improve product purity, decrease batch rejection, and increase process performance. Systems involving the production of one crystalline solid while suppressing the nucleation of a second solid are common, with applications to polymorphs, hydrates/solvates, chiral resolution, and byproducts. Several process analytical technologies (PATs) have already been established for system-specific detection of an undesired solid phase; this study adds to that PAT suite an image-based technique with greater generality and sensitivity than common tools such as Raman spectroscopy and focused beam reflectance measurement. In situ microscope images are analyzed with a convolutional neural network (CNN) to extract image features and classify cropped regions obtained by a sliding window as containing a single particle type or multiple particle types. As an experimental case study, the performance of the technique is evaluated using a system involving contamination of reactive crystallization of cephalexin with phenylglycine, the sparingly soluble byproduct in the enzymatic synthesis of cephalexin. A CNN, ResNet, was retrained for the classification task at hand and showed >98% accuracy on the test data, highlighting the distinct features of different crystal classes used as the basis of process monitoring.