A synthetic machine learning framework for complex crystallization processes: The case study of the second-order asymmetric transformation of enantiomers

Abstract

Improving the flexibility and capabilities of pharmaceutical technologies and lowering the process development time while satisfying regulatory constraints translate to highly desired resilient supply chains. Generating and validating the necessary knowledge in-silico before experimentation or plant-scale intervention with minimized human implications is the dream this paper aims to contribute to. The idea is demonstrated through the case study of particle size-controlled second-order asymmetric transformation of enantiomers (SOAT), a crystallization and racemization-based resolution technique. The fundamental processes form a complex network, to which we propose a crystallization-integrated wet milling system, which is promising but has limited general, and no SOAT-specific operation experience. The developed computational framework has three steps: (1) skilled scientists draft potential technologies to solve a well-defined problem; (2) all alternatives are modeled with high fidelity first-principle models, followed by high throughput parametric optimizations (3) data mining is deployed to extract key information from the synthetic database of optimal solutions and take the effectiveness of process design and operation to the next level. In this work, population balance-based models were applied to describe the crystallization, fragmentation, and chemical reaction in the crystallizer and wet mill. The efficient implementation enabled the execution of nearly 1300 global multiobjective process optimizations. Carefully trained classifiers consistently predicted the with or without milling and with or without T-cycling decisions (AUC score exceeding 0.85). Regression methods estimated the product property profile with Pearson coefficients over 0.95. The results agree with the existing SOAT literature based on a literature meta-analysis, but some of the conclusions point beyond the state-of-the-art.