Optimierung des Betriebs von Wirbelschichtverfahren mittels maschinellen Lernens

Fluidized beds are the basis for scores of applications in which fast mixing, heat and mass transfer of gas and solid particles are essential. Their performance largely relies on the bubble dynamics: rising bubbles drive the solids circulation and significantly enhance gas-solids contact, improving mixing, reactions, and transport properties. So far, almost all fluidized beds are operated with a uniform gas flow. However, some recent academic work shows that operating a fluidized bed with an alternating gas flow (e.g. sinusoidal gas fluidisation velocity) leads to different bubble patterns and dynamics. In this project, we aim to control the bubbles in a fluidized bed, by application of computational intelligence (CI) methodologies such as evolutionary algorithms and genetic programming. We will use our lab-scale fluidized bed with camera system and our model developments in the Eulerian-Eulerian and Eulerian-Lagrangian frameworks to capture the dynamics of bubbles in the fluidized bed as the fluidizing gas velocity is spatio-temporally varied. Firstly, these results will be used to find the optimal inflow-pattern for given target functions. The challenge for the CI algorithm is to find the right balance between the computationally and timely intensive experimental data and the simulation data to efficiently deliver the required fluidization velocity profile. In addition, we aim to address multiple conflicting target functions using multi-objective optimization algorithms. Secondly, the CI algorithm will be used to steer and control the velocity profile, to obtain a specified bubble size and dynamics. Being able to control the behavior of the bubbles in a fluidized bed will significantly improve the desired outcome, such as product quality, efficiency and selectivity of the process, to name a few.