(545a) Neural Networks-Based Self-Driven Control Strategies for a Pharmaceutical Crystallization Process

In the pharmaceutical manufacturing, crystallization is the most applied downstream process to recover high-purity pharmaceuticals and fine chemicals. However, establishing control over this unit operation is still a major bottleneck particularly in the industry. The main reason is the nonlinear and stochastic dynamics of the crystallization process that challenge the control strategies and actions. In addition to this, in the absence of any online or inline measurements, the nonlinear relationship between the process outcome (e.g. crystal properties) and process inputs is difficult to understand and to establish. Fortunately, the latest advances in Process Analytical Technology (PAT) tools enable to monitor the critical solution- and crystal-state properties in real-time. This eases the understanding of the crystallization dynamics and helps to establish a correlation between the operating inputs and the critical properties of the crystal product. Thus different and advanced feedback control strategies can be applied based on the information obtained from PAT tools in order to establish some degree of control over crystal size, polymorphic form and morphology. Despite the significant progress, the control of industrial crystallization process is still an outstanding issue. One reason behind might be attributed to the fact that the state of the art crystallization control strategies fall into one of two restrictive categories: model-free (direct nucleation control, supersaturation control) and model-based. While model-free methods have the advantage of being simple and widely applied, some efforts are needed for the determination of a robust setpoint or robust sensor calibration that should be also valid in the presence of the disturbances in the system. On the other hand, model-based control strategies suffer from the complexity and the accuracy of the implemented model as well as the speed of solution time in order to achieve a reliable real-time optimization. However, a fast and successful transition between different scales of the crystallization operation in the pharmaceutical industry requires a simple but reliable control strategy. Therefore, differently from two categories, a third category can be attributed to the data-driven (based on process data) control strategies. Data-driven approaches have proven to be useful particularly for the complex processes that are troublesome and costly to develop a knowledge-driven model in a timely manner for the optimization and control of industrial processes. Therefore, in this work we present a data-driven control strategy based on the neural networks (e.g. radial basis functions and deep neural networks) for a pharmaceutical batch cooling crystallization process. The proposed data-driven control utilizes both real-time and offline the experimental data measured with different PAT tools (e.g. FBRM, FT-IR, image analysis) to optimize the cooling profile in order to achieve the desired crystal-state property profile throughout the process. A model pharmaceutical compound (e.g. ibuprofen) is selected as a case study and the robustness of the proposed control strategy is shown based on the comprehensive experiments in the presence of several process disturbances (initial supersaturation, impeller speed, the impurity of the solvent and seed size). The proposed neural-networks based control is a promising strategy that is easy to implement, fully-automated, and relies on relatively limited data for training. Hence, it is expected to contribute to quick process development and control, especially in the absence of comprehensive process understanding and historical data in pharmaceutical industries.

Acknowledgment

We would like to thank the Danish Council for Independent Research (DFF) for financing the project with grant ID: DFF-6111600077B.