(627d) Enhancing Crystallization Process Design Using Bayesian Optimization

Bayesian optimization (BO) has recently emerged as a powerful methodology in reaction engineering and pharmaceutical process development, enabling rapid optimization of experimental conditions while minimizing resource utilization. When coupled with automated parallel experimentation platforms, BO has immense potential to drive the shift toward intelligent, autonomous pharmaceutical process design.¹ However, despite its widespread application in reaction engineering, the use of BO within pharmaceutical crystallization process design remains relatively unexplored.

In this study, we propose and demonstrate a novel, systematic integration framework that incorporates BO methodologies directly into existing experimental workflows for pharmaceutical crystallization. We evaluate the capability of BO to sequentially and in parallel design experiments that yield crystals with desired critical quality attributes (CQAs) such as yield and crystal size distribution. This evaluation is initially performed in-silico, utilizing a mechanistic population balance model as ground-truth to objectively compare BO-suggested solutions with those obtained from mechanistic population-balance models. The influence of measurement uncertainties (noise) on the BO-driven experimental designs was further investigated, providing insights into the robustness of these methods under realistic laboratory conditions.

Comprehensive analyses across multiple crystallization scenarios—ranging from single-objective and multi-objective to high-dimensional and constrained BO formulations—were conducted, revealing both the strengths and limitations of existing BO approaches. Insights derived from these analyses clarify key implementation challenges and inform the adaptations required to seamlessly integrate BO methodologies within conventional crystallization process development practices. To extend BO’s applicability, we explore its use in optimizing dynamic input trajectories such as temperature and antisolvent additional profiles. To do so, variational autoencoder networks (VAEs) were used to project these trajectories into a lower-dimensional latent space, making them amenable for Bayesian optimization.²

The proposed approach is demonstrated through an industrial case study on the crystallization of Compound X, wherein multiple thermocycle temperature profiles and their effects on crystal size distribution and degree of agglomeration were systematically explored using BO.

Acknowledgement:

Funding from Takeda Pharmaceuticals International Co. is gratefully acknowledged.

References:

1. Chang, H., Domagalski, N., Tabora, J. E. & Tom, J. W. Bayesian data-driven models for pharmaceutical process development. Curr Opin Chem Eng 45, 101034 (2024).
2. Valleti, M., Vasudevan, R. K., Ziatdinov, M. A. & Kalinin, S. V. Bayesian optimization in continuous spaces via virtual process embeddings. Digital Discovery 1, 910–925 (2022).