(39g) Developing Digital Twins for Pharmaceutical Crystallization Processes Using Machine Learning-Based Strategies

Population-balance-based mechanistic digital twin development has seen increasing adoption within the pharmaceutical industry, facilitating the design of crystallization processes to ensure consistent delivery of crystals with desired attributes such as purity, crystal size distribution (CSD), yield, and polymorphic form. However, accurately developing system-specific population balance models (PBM) presents challenges, including the identification of suitable semi-empirical equations to describe various mechanisms and the estimation of kinetic parameters from experimental data.¹ Moreover, the high computational cost of first principles PBM models hinders their real-time application, whether for updating crystallization kinetics based on new process data or for real-time optimal control under uncertainty. To this end, data-driven digital twin development using state-of-the-art machine learning (ML) and artificial intelligence (AI) methods presents a promising avenue for providing a viable modeling alternative, overcoming the challenges associated with mechanistic digital model development for complex pharmaceutical systems, and in some cases, even surpassing the capabilities of mechanistic digital twins.²

While a notable surge in literature demonstrates proof-of-concept studies applying data-driven strategies for modeling and controlling crystallization processes, a significant research gap persists, particularly concerning pharmaceutical APIs. This gap is characterized by limited experimental connections, either through training models with experimental data or validating models against experimental results.² Thus, the primary objective of this research is to address this gap by devising systematic frameworks for constructing data-driven digital twins from experimental data and assessing their effectiveness for robust design and control of pharmaceutical crystallization processes.

The framework development was initiated using a model compound from the literature with a needle-shaped morphology described using a complex two-dimensional PBM model.³ The data-driven digital twin development was conducted in phases mirroring a typical pharmaceutical API development procedure followed in the industry: small-scale Crystalline experiments, lab-scale crystallization experiments, and pilot-scale experiments. For each phase, the training data used for building the data-driven models was constructed using specific experimental information available during each phase. For example, only the operating conditions and crystal quality attribute information (yield and CSD in both length and width directions) were used in the first phase, while the second stage additionally consisted of calibrated process analytical technology (PAT) data from experiments done at lab-scale. The machine learning development workflow included key steps such as data pre-processing, evaluation of different model architectures, optimization of various hyper-parameter settings, and model testing. To address the commonly observed training data-scarcity challenge associated with developing data-driven digital twins from experimental data, state-of-the-art deep learning-based generative modeling techniques were evaluated and used for generating synthetic experimental data from real data. Furthermore, augmentation of real and synthetic data was performed to enhance the training process of neural networks, thereby improving their prediction capabilities. Lastly, the developed data-driven digital twin framework was applied and validated using the experimental data from a real-life API crystallization process.

Acknowledgement:

Funding from Takeda Pharmaceuticals International Co. is gratefully acknowledged.

References:

1. Nagy, Z. K. & Braatz, R. D. Advances and new directions in crystallization control. Annu. Rev. Chem. Biomol. Eng. 3, 55–75 (2012).
2. Xiouras, C., Cameli, F., Quilló, G. L., Kavousanakis, M. E., Vlachos, D. G. & Stefanidis, G. D. Applications of Artificial Intelligence and Machine Learning Algorithms to Crystallization. Chem. Rev. 122, 13006–13042 (2022).
3. Wu, W. L., Chappelow, C., Hanspal, N., Larsen, P. A., Patton, J., Shinkle, A. & Nagy, Z. K. Digital Design of an Agrochemical Crystallization Process via Two-Dimensional Population Balance Modeling. Org. Process Res. Dev. 28, 558 (2023).