Project, Bachelor and Master theses

Fast alle Strömungen, die in der Umwelt und Technik vorkommen, sind turbulent. Jedoch ist bereits die numerische Simulation einphasiger turbulenter Strömungen aufwendig, wobei es viele erfolgreiche Modellierungsansätze gibt. Eine noch größere Herausforderung besteht hingegen in der numerischen Analyse partikelbeladener turbulenter Strömungen. Trotz ihrer hohen Relevanz in Umwelt und Technik, sind vorhandene Modelle nur für vereinfachte Bedingungen gültig und eine Validierung steht oft noch aus. Ein wichtiger Anwendungsfall ist die numerische Auslegung einer Biomasse-Brennkammer. Dabei ist die Bestimmung der Aufheizraten, der Dynamik, und der turbulenten Durchmischung nicht-sphärischer Partikel entscheidend um den gesamten Verbrennungsprozess zuverlässig auszulegen. Die Generierung von hoch-aufgelösten Referenzdaten mit Hilfe von Simulationen und die Entwicklung von genauen Modellen für Anwender, sowie deren Validierung sind aktuelle Forschungsvorhaben, die am Aerodynamischen Institut intensiv verfolgt werden. Für dieses Projekt sind wir auf der Suche nach motivierten Masterarbeitern.

Particle laden turbulent flows play an import- ant role in many technical fields such as drug de- position in human airways, or the combustion of solid biofuels. To simulate these problems, two dif- ferent approaches are available: The fully resol- ved direct numerical simulation (DNS), where all features of the flow field including the surface of the particles are resolved, and the Lagrangian point-particle approach, where empirical models are used to determine the forces acting on each particle. Due to the high computational costs of the DNS approach, reduced order point-particle models are indispensable for the simulation of lar- ge scale technical applications. These models pro- duce good results for very small particles but lead to significant errors for larger ones. In this thesis, data from a DNS will be used to train an artificial neural network (ANN) which predicts the forces acting on a particle in a reduced order point-particle model. In a first step, fully re- solved simulations with many particles have to be performed to extract the local flow field and force data for each particle, resulting in a large databa- se for the determination of new point-particle mo- dels. In the second stage, the ANN is trained based on this database, where velocity data from a par- ticle’s surrounding flow field function as input, the acting force as output, and the acting forces from the DNS as ground truth. In a final step, it is inve- stigated how transfer learning improves the pre- dictive capabilities of the ANN. That is, the ANN is pre-trained with ground truth data from empi- rical drag laws, and the highly resolved data from the DNS are employed solely in the final training iterations.

Hydrogen combustion is shaping to be an im- portant component for the future of electricity grids based on renewable energies. Combustion si- mulations solve the mass, momentum and energy equations in addition to equations describing the reaction kinetics and transport of species. The- se additional equations are, in general, expensive to solve when compared to the non-reacting flow equations. In order to reduce computational expense, some simplified alternatives exist, like skeletal reaction mechanisms or two-step chemistry. If one desires to use the complete, relevant reaction chain, tabu- lated chemistry or machine learning models are a good alternative and a good compromise between accuracy and efficiency. In this context, four ty- pes of machine learning algorithms are employed. The first algorithm maps each location in the high- dimensional flow domain to a low-dimensional sub- space. With the second algorithm, representations in the low-dimensional subspace are clustered into regions of different thermochemical equilibrium. A classifier is then trained to classify new points into You ... these regions. Finally, reaction kinetics and trans- port species are learned for each cluster by artificial neural networks. The main tasks of the thesis are to (1) im- plement the proposed machine learning-based me- thod, and (2) compare the accuracy and speed of the proposed method compared to solving the re- acting flow equations, chemical kinetics and trans- port phenomena directly.