Seminar on "Application of the Koopman operator and machine-learning for data-driven modelling of particulate processes"

Abstract: Particulate materials and particulate processes are of great importance for chemical, pharmaceutical and food industries. A typical example is the crystallization of the final product from a solution or as part of a purification step. In general, product properties are directly connected to the particle properties. For example, the flowing properties depend on the particle size distribution and the particle moisture content. As these processes are often subject to great uncertainties and may exhibit unstable behavior there is a great need for control in order to guarantee a constant product quality. In this talk, the focus will be on fluidized bed spray granulation and its control. Here, model-based and data-driven control approaches will be presented.

Traditional model-based approaches for complex systems like particulate processes are often hindered by simplifying assumptions, resulting in inaccurate predictions and suboptimal control strategies. Moreover, these methods frequently neglect the valuable data available to improve model accuracy.  To overcome this limitations, the Koopman operator framework is established, which transforms nonlinear dynamical systems into a linear representation in Hilbert space. This enables the extraction of meaningful features from complex datasets.

Utilizing Koopman-based numerical techniques such as Dynamical Mode Decomposition (DMD), it becomes possible to directly compute the linearized system from data without requiring explicit physical models. The resulting models allow for the application of well-established control and analysis methods for linear systems. In this presentation, we demonstrate the practical application of Koopman operator-based modeling on the example of particulate processes.

Sponsors: