Deep Learning-Based Oriented Object Detection for In-Situ Image Monitoring and Analysis: A Process Analytical Technology (PAT) Application for Taurine Crystallization

Abstract Image analysis enables the estimation of critical process properties such as crystal size, morphology, and crystallization kinetics. Despite the rich image information, the lack of a robust image analysis technique has been an obstacle to promote its applications. In this work, an automated image analysis technique that combines the state-of-the-art oriented object detection model, S2A-Net, was developed for in situ estimation of the two-dimensional crystal size distribution (CSD) and the crystal counts. The model was trained to detect and classify both crystals and clusters to enable quantification of the extent of agglomeration and exclude unreliable detections. The effectiveness and robustness of extracting size and aspect ratio at various image complexities were verified by comparing with the focused beam reflective measurement (FBRM) and manually analyzed images in the experimental studies for taurine batch cooling crystallization with different seed loadings. An on-line calibration strategy for solute concentration measurement with Raman spectroscopy was introduced to eliminate the dedicated calibration experiments. The secondary nucleation and growth rate kinetics were evaluated from the on-line measurements and validated by numerical simulation. The proposed method provides a novel PAT strategy that enables accurate two-dimensional size measurement and shape characterization for online monitoring and control of a solution crystallization process.