This project is to be conducted in cooperation with TU Dortmund (Prof. Markus Thommes).

The overall aim of this project is to develop and implement a data-driven control algorithm for autonomous powder compaction on a rotary tablet press. This includes an online autonomous adjustment of quality attributes such as dose and hardness, while minimizing the waste during start-up, maximizing the production rate and compensating the process disturbances during production. Our interpretation of the autonomous process behavior implies an online and adequate self-adaptation of the setpoints of the process parameters such as punch distance, impeller speed, etc. in attempting to maintain the product quality. The latter may indeed undergo various deteriorations as a result of the impact of inherent process disturbances, such as material flow fluctuations or varying material properties (e.g. particle size). In combination with a profound process understanding, this leads to optimized individual process steps (feeding, blending, flling, compression, ejection; see below). The objective thereby consists in enhancing and improving the product quality and process effciency in comparison to the manual process management, which has been a common practice in the pharmaceutical industry today. To this end, we develop offline and online data-driven predictive control (DPC) policies utilizing the behavioral system theory. Initially, a theoretical base for the decision policies subject to stochastic and nonlinear data-based process models needs to be established in both scenarios, offline and online adaptation, which go beyond the state-of-the-art of data-driven control. The data-driven approach boils down to formulations and solutions of optimal control problems (OCP) on the basis of data-driven modeling schemes. In this project, we consider a rather general class of polynomial NARX models due to their powerful description capabilities, their relevance in practical applications and the proven control feasibility in preliminary studies. The online adaptation requires additionally a suitable adaptation law of the data Hankel matrices based on the process observations.