Student Theses

Combustion processes are central to a wide range of industrial applications, from energy conversion in internal combustion engines to high-value material production such as nanoparticle synthesis in pharmaceuticals and cosmetics. Despite their importance, combustion systems remain extremely challenging to model due to the strong coupling of turbulent transport and highly nonlinear chemical kinetics. The wide separation of spatial and temporal scales – from molecular reactions occurring in nanoseconds to flow structures persisting over milliseconds – makes direct numerical simulation infeasible for most practical cases. Stochastic particle methods offer a promising framework for resolving the detailed chemical processes underlying combustion. However, the efficiency and accuracy of these methods critically rely on the particle number and density. To ensure a sufficient number of stochastic particles, the particle population has to be actively managed during the simulation, adding or deleting particles where needed. Investigating and optimizing such particle management techniques is therefore essential for advancing predictive combustion modeling and enabling more sustainable and efficient energy technologies.

Super-resolution (SR) reconstruction of turbulent flows aims to generate high-resolution flow fields from low-resolution data that are more readily available from numerical simulations and experimental measurements. Figure 1 shows an example: the model SwinFIR-FT largely reconstructs a two-dimensional flow field, which is sampled from a direct numerical simulation (DNS), from its low-resolution filtered counterpart. Deep neural networks have achieved notable success as SR models. Many network architectures and loss functions have been proposed. An objective is to train the model with flow samples at low Reynolds numbers and subsequently generalize it to reconstruct turbulent flows at high Reynolds numbers, which are prohibitively expensive to simulate or capture experimentally. However, to date, the model’s generalization performance remains dependent on the input grid resolution (i.e., ratio of grid size to Kolmogorov length scale), which is difficult to estimate accurately for low-resolution data in practice. There are workarounds that train the model using samples featuring multiple grid resolutions or add multiple upsampling branches to the model architecture. Nonetheless, the generalization flexibility is still limited with these workarounds. Therefore, advanced training strategies/model architectures are being sought to remove the dependence of generalization on grid resolution.

The rapid advancement of international space transportation has created the need for innovative, lightweight rocket engines capable of multiple restarts, while still meeting modern environmental standards. One promising approach combines the combustion of cryogenic propellants with laser-based ignition. Before ignition, however, the cryogenic liquid is injected into the combustion chamber at pressures far below the saturation point, causing instantaneous boiling—a phenomenon known as flash evaporation. The flow behavior and physical processes at the injector exit and face plate remain insufficiently characterized, emphasizing the necessity for further numerical investigation.

Nanoparticles are of great relevance in various fields, e.g. filtration/porous media, groundwater flows, atmospheric science and combustion processes, as they are difficult to capture and collect due to their small size.  Therefore, nanoparticle collection is subject to ongoing research and requires detailed knowledge of their movement and interactions with the flow field and surroundings, which is generally referred to as nanoparticle dynamics. Numerical simulations can be used to precisely calculate the nanoparticle dynamics and improve their collection efficiency.

Hydrogen, increasingly popular as an alternative fuel, presents unique challenges due to its high reactivity and associated ignition and explosion hazards. With the expected rise of hydrogen production and use in both public and industrial locations, understanding and predicting the consequences of gas explosion are crucial for ensuring safety. Hydrogen is stored in tanks in liquid or gaseous form under high pressure. If the pressure in the tank rises due to any technical failure, some hydrogen must be released to prevent the tank from rupturing. In this project, fully resolved simulations of canonical problems such as flame propagation in channels with obstacles will be performed to inform model development that may be used for industrially relevant scales to predict under which conditions the released hydrogen can lead to detonations. This project is done in collaboration with Shell.

The recent rapid development of international space transportation demands novel, lightweight rocket engines with restart capabilities, while fulfilling current environmental regulations. The combination of combustion of cryogenic propellants with laser ignition is a promising solution. However, prior to ignition the cryogenic liquids are injected into conditions far below their saturation pressure leading to immediate boiling, so called flash evaporation. The flow field and processes at the injector exit and face plate are not yet fully understood and further research is required.

Reliable simulations pertaining to soot remain a considerable challenge. One of the primary reasons for this challenge is the intricate nature of fuel/soot mechanisms, which can consist of hundreds of species. Direct integration (DI) of the extensive and stiff ordinary differential equations (ODEs) accounts for over 90% of the total CPU hours, thereby limiting the feasibility of high-fidelity soot simulations. Artificial neural networks (ANNs) are a powerful tool for emulating non-linear systems. They are universal function approximators and have both the advantages of high mapping complexity and high computational speed. Therefore, utilizing ANN techniques for modeling thermodynamic evolution in sooting flames holds great promise.

Bioreactors are an important part of many industries. They are used in medical applications for the production of vaccine components, in the food industry—like beer brewing—the production of biogas for providing sustainable energy or the treatment of sewage and waste recycling. They work by having microorganisms or enzymes submerged in a liquid, which is continually provided with oxygen, nutrients and reactants. To better understand the processes inside bioreactors and to optimize current reactor designs, simulations of the conversion inside the reactors can be performed. However, the complex physical processes and the different time scales governing industrial-scale reactors make detailed simulations with current computing power difficult.

Agglomeration of small particles is an important growth mechanism in many industrial and natural processes. It plays an essential role during the formation of soot and the production of flame-made commodities like fumed silica or titania which are used, for example, as agents in the food-processing and pharmaceutical industry. As the shape and size of the forming agglomerates determine the product characteristics, it is desirable to develop suitable models that help predict and control the growth dynamics in these processes.

Agglomeration is widely applicable across a variety of industrial fields, such as pharmaceuticals, materials science, and biotechnology, playing a significant role in enhancing both operational efficiency and product quality. However, most research has primarily concentrated on agglomerating uniform-sized primary particles, with limited attention given to polydisperse primary particle size distributions. It is essential to develop models for understanding the growth dynamics of agglomerates composed of polydisperse primary particle size distributions, as this can ultimately contribute to the enhancement of the final product’s quality.

The pursuit of clean energy sources requires the search for novel methods for generating, storing and transporting these. Among them, hydrogen has emerged as a solution capable of addressing society’s growing demands and the call for more environmentally friendly substitutes. However, accurately and efficiently simulating hydrogen combustion, as well as appropriately modeling its underlying physical phenomena, remains a significant challenge.