Harnessing Birefringence for Real-Time Classification of Molecular Crystals Using Dynamic Polarized Light Microscopy, Microfluidics, and Machine Learning

Abstract

Molecular crystals are ubiquitous in a variety of industrial contexts, from foods to chemicals and pharmaceuticals. The timely identification of different molecular crystal forms (and transformations between forms) is critical in both manufacturing and chemical/pharmaceutical product design, as they possess different physicochemical properties (e.g., solubility, melting and boiling point, etc.) that could affect product attributes such as stability and dissolution rate. Current characterization methods typically involve a time delay between sampling and analysis and are unable to directly quantify forms/transformations in crystal ensembles at a single crystal level in real time. Here, we introduce a new methodology to accomplish such measurements, which utilizes a combination of microfluidic flow cells, machine learning, and a rotating polarizer−analyzer pair with orthogonally aligned polarization axes for imaging and automated access to interference colors of birefringent molecular crystals that are characteristic of the polymorphic form. Since the polarized light microscopy images of the crystal ensembles captured represent their instantaneous states at the time of acquisition, the methodology uniquely enables real-time, in situ quantification of polymorphically mixed pharmaceutical crystals in both static (polymorph or pseudopolymorph mixtures) and dynamic crystallization systems (e.g., solution mediated phase transformations). The classification of crystal ensembles (∼3000 crystals classified in under 10 s) at a single crystal level can be achieved with an accuracy of ∼86% (azithromycin dihydrate and azithromycin sesquihydrate) to 94% (α-glycine and βglycine). This sheds quantitative insights into the dominant crystallization phenomena such as nucleation, growth, or dissolution, potentially enabling both process monitoring as well as extraction of crucial kinetics data needed for crystallization process modeling and control. We envision the applicability of this methodology in accelerating the exploration of storage, process condition, or additive dependent polymorphic form outcomes that are of interest during early stage research and development when limited quantities of materials are available.

Recent Advances in the Application of Machine Learning to Crystal Behavior and Crystallization Process Control

“…Tailoring crystals for desired habits are crucial in various applications, but traditional methods often rely on a combination of experience and trial-and-error. These approaches can be time-consuming, resource-intensive, and often lack predictability, hindering efficient material development …”