(259d) Data-Driven CLD-to-Psd Model for in-Line Crystal Size Monitoring

Recently, a data-driven model architecture was proposed for the correlation of CLD to PSD data [1,2]. The model quantifies the particle size distribution (PSD) based on solids’ concentration data and the corresponding chord length distributions (CLD) measured inline using FBRM [1] and has been shown to be very robust at estimating various particle morphologies, concentrations and size ranges typically encountered in pharmaceutical crystallization processes. The model architecture consists of three distinct steps: first, the CLD data is compressed into moments, or singular values that describe the entire distribution. Second, advanced regression models convert these moments into a small set of PSD percentiles (also moments). Finally, these PSD percentiles are mapped into an alternate composite distribution using a set of parametrized functions called generating functions [1]. In this work, further progress on steps 2 and 3 is reported which is aimed at increasing model robustness, particularly when only a limited amount of data is available. Furthermore, model performance has been assessed for process monitoring under real-life conditions, including multi-modal size distributions. The technical issues of dataset generation and process monitoring are discussed as well.

[1] Irizarry R, Chen A, Crawford R, Codan L, Schoell J. Data-driven model and model paradigm to predict 1D and 2D particle size distribution from measured Chord-length distribution. Chem Eng Sci. 2017; 164: 202-218.

[2] Schoell J, Irizarry R , Sirota E, Mengel C, Codan L, Cote A. Determining particle size distributions from chord length measurements for different particle morphologies. AIChE J. Vol 65.