(577c) A Comparative Study of Surrogate Model Based Control Strategy for a Pharmaceutical Crystallization Process

Authors: 
Öner, M., Technical University of Denmark
Pal, K., Purdue University
Montes, F., Technical University of Denmark
Stocks, S. M., LEO Pharma A/S
Abildskov, J., Technical University of Denmark
Nagy, Z. K., Purdue University
Sin, G., Technical University of Denmark
Abstract

In pharmaceutical manufacturing, crystallization is the most preferred technique in downstream processing to recover crystal products. Crystal quality attributes such as crystal size distribution, purity, and polymorphic form, which are critical for not only following unit operations, but also therapeutic properties of the formulated product, are determined by the operating conditions of crystallizers. Therefore, control of a crystallization process is critically and essentially the control of quality attributes. Traditionally, crystallization processes are operated based on a pre-defined recipe of a process variable such as a temperature trajectory for a cooling crystallization system, and the quality of the crystal product is tested only at the end of the process. Quality by testing (QbT) approach leads to high profit loss since batch-to-batch variations of crystal quality attributes are unavoidable, due to the absence of control actions taken during the process in order to avoid effects of any disturbances available in the system through e.g. feedback signals coming from solid-state attributes or supersaturation. However, consistency of a process outcome is highly dependent on a robust process control strategy. In literature, emergence of the latest generation of process analytical technology (PAT) tools, mechanistic understanding and modeling of underlying mechanism of crystallization process and incentives by regulatory mechanism have accelerated the efforts on the development of advanced process control strategies in order to achieve desired product quality, so called quality by control (QbC). State of the art crystallization control strategies can be classified into two main groups as model free/data based and model based. As a model free method, direct nucleation control (DNC) is based on the principle of operating at a predefined set point of crystal number density to dominate crystal growth, while supersaturation control (SSC) as another model free method that aims to keep supersaturation level within metastable zone and constant over the experiment time. While the model free methods are simple and found many applications, they might require some investigations on the determination of a robust set point as well as some efforts on online sensor calibration. Model based control strategies utilize real time crystallization process simulations to predict the effects of inputs and disturbances on the process output, whose performance is strictly dependent on the complexity and accuracy of the implemented model as well as speed of solution time. However, often a simple but reliable control strategy is required in the pharmaceutical industries to achieve a fast and successful transition from laboratory to pilot scale crystallization operation. Therefore, in this work we aim to benefit from the simplicity of data based models, while maintaining the accuracy of the predictions in an acceptable range. To this end, we present a surrogate model based control strategy for a pharmaceutical batch cooling crystallization process. Two surrogate models namely Radial Basis Functions Network (RBF) and Polynomical Chaos Expansion (PCE) are utilized to follow a reference trajectory of a mean crystal size by manipulating and updating the temperature profile during process simulation. Performance of the proposed control strategy based on two surrogate models of RBF and PCE is evaluated in the presence of disturbances in the system such as initial concentration, temperature and seed specifications. Additionally, as a proof of concept, experimental applications of the proposed methods are supplemented.

Acknowledgment

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