(660g) Automated Continuous Crystallization of APIs with Real-Time Crystal Size Analysis Via Laser Diffraction and Closed-Loop Optimization By Machine Learning

A state-of-the-art autonomous experimental platform was developed to facilitate the process development of continuous antisolvent crystallization of micron-sized crystals of active pharmaceutical ingredients (APIs). This platform is integrated with online laser diffraction technology, enabling the real-time measurement of crystal size distributions (CSDs). A confined impinging jet reactor was used as antisolvent crystallizer, offering superior mixing and yielding consistent CSDs that outperformed those obtained from batch crystallizers. Antisolvent crystallization of APIs can be tuned by manipulating the various input parameters of the process, but identifying optimal conditions within the multivariable design space to target a specific mean crystal size poses a significant challenge. To reduce the experimental effort and navigate the design space efficiently, powerful tools such as design of experiments (DoE) and online CSD measurements are essential, because, coupled with automated real-time process control, these can significantly expedite the exploration of the design space. Usually, crystal suspensions must be suitably diluted to avoid errors from multiple light scattering, which is the key challenge when online laser diffraction is implemented. This work overcomes these obstacles by automatically collecting and diluting the samples that elute from the crystallizer in a collection vessel, prior to flowing it to the optical flow cell of the laser diffraction analyser. A comprehensive integration of various software tools, namely LabVIEW, Python and PharmaMV, along with logic algorithms, data analysis and post processing has been implemented within the platform. This integration facilitates the automated control of all sensors and equipment, thereby ensuring seamless and fully automated operation.

However, initial experiments with the automated platform resulted in clogging. To address this issue, classification models were used for the identification of the feasible operating space for the crystallization process. This becomes particularly critical in situations where there is limited prior information available, a common occurrence during the early phases of API development. Once the feasible operating space was determined, the next step in our methodology involved DoE methods to generate initial experimental conditions in the feasible space and conducting automated experiments within this space to train and calibrate machine learning (ML) regression models. Multi-task Gaussian process models and active learning were implemented for surrogate modeling and optimal sequential experimental design, respectively. Finally, Bayesian optimization methods were implemeted in the platform for optimization of particle size, to create closed-loop self-optimizing platforms for API crystal manufacturing. A tailored graphical user interface was developed for autonomously operating the crystallization platform, conducting DoE-based experiments within specified parametric ranges, and visualizing key metrics such as mean crystal size and crystal size distribution of crystal suspensions, as well as the progression of model development and of the crystal size optimization stages.

The autonomous crystallization platform rapidly screens the process parameters and identifies the optimal conditions for achieving the desired size of API crystals through the utilization of machine learning algorithms. The platform has been showcased to produce 2-80 mm sized ibuprofen and ketoprofen crystals.