Recent Submissions
Study on constraining turbulence to met mast data from the WINSENT complex terrain test site for use as inflow for CFD simulations
(2025) Müller, Carsten
This work presents a method for generating time-resolved, three-dimensional turbulent inflow conditions for URANS/DDES simulations using the flow solver FLOWer. Turbulent inflow fields are generated using the Mann Turbulence Model via python’s Hipersim package and are applied as boundary conditions in the solver. The inflow is to reproduce both the absolute values and spectral characteristics of single-point time series and 3D velocity fields represented in atmospheric turbulence. A dedicated toolchain, InFlow, was developed to process and adapt turbulence input data from the WINSENT test site near Stötten, Germany. The approach is designed to be computationally efficient, straightforward to apply, and accurate enough for use in practical wind energy simulations. Its performance and limitations are evaluated across varying inflow scenarios and setups.
SwinDiffuser : accelerating diffusion models through parallel processing
(2024) Ye, Yun
Diffusion models have emerged as a powerful generative approach in artificial intelligence, particularly for image, video, and audio synthesis. Despite their success, these models suffer from significant computational demands due to the iterative nature of the denoising process. This thesis introduces the SwinDiffuser, a novel method designed to accelerate diffusion models by leveraging parallel processing. The proposed method divides high-resolution images into smaller patches, allowing for simultaneous processing by multiple diffusers. Key innovations include the integration of global feature extractors and shifting windows to maintain coherence across patches, and the utilisation of a U-Net architecture for noise prediction. Experimental results demonstrate that the SwinDiffuser achieves comparable image quality to standard diffusion models while significantly reducing generation time. This advancement paves the way for practical applications of diffusion models in real-time scenarios and resource-constrained environments.
Data attribution for diffusion models
(2024) Bien, Tanja
Diffusion models have demonstrated a remarkable ability to generate photorealistic images. However, it is difficult to explain what causes the generated image. Tracing the output back to the training data and identifying the most influential samples is necessary to debug the model, find biases, or provide fair compensation to creators. While data attribution methods have been extensively studied in the supervised setting, data attribution for generative models such as diffusion models remains a challenge. The aim of this thesis is to provide an overview of existing methods for data attribution and evaluation methods. In the absence of a commonly used benchmark, a framework for evaluating data attribution methods was implemented as part of this thesis. Various experiments and evaluation methods allow a comparison between the different methods to better understand their use cases and limitations. Furthermore they lead to the proposal of new normalization method, called loss-normalization.
Stochastic query synthesis for neural PDE solvers
(2025) Ullrich, Finn
PDEs are highly influential in physics and are describing various phenomena in the world, from wave movement to electro-magnetics. The problem arises when one tries to solve them, which requires enormous computing power for a numerical solution. To overcome the limitiations, neural PDE solvers have been proposed, using neural networks to approximate the solution trajectories. However, neural PDE solvers require training data from an computationally expensive numerical solver. Therefore, Musekamp et al. created a benchmark, which investigates active learning for neural PDE solvers. Active learning can reduce the amount of data required, while keeping the same performance. In this work, we will demonstrate a new strategy of selecting samples called stochastic query synthesis. Following this, we will remove the pool currently used and rather create a Markov chain directly from the input space containing unlabeled instances. The transition probability is based on the unadjusted Langevin algorithm, allowing us to sample by exploiting gradient information. To retrieve a better result, instead of just one chain, we will create multiple parallel chains, and only take the last state as input. We will show that this approach is equally effective as the currently implemented pool-based implementation. However, there are still performance problems that need to be solved in the future, to make it viable in practice.
Model-based reinforcement learning under sparse rewards
(2023) Akash, Ravi
Reinforcement Learning (RL) has recently seen significant advances over the last decade in simulated and controlled environments. RL has shown impressive results in difficult decision-making problems such as playing video games or controlling robot arms, especially in industrial applications where most methods require many interactions with the system in order to achieve good performance, which can be costly and time-consuming. Model-Based Reinforcement Learning (MBRL) promises to close this gap by leveraging learned environment models and using them for data generation and/or planning and, at the same time trying to be sample efficient. However, Learning with sparse rewards remains a significant challenge in the field of RL. In order to promote efficient learning the sparsity of rewards must be addressed. This thesis work tries to study individual components of MBRL algorithms under sparse reward settings and investigate different design choices made to measure the impact on learning efficiency. Suitable Integral Probability Metrics (IPM) are introduced to understand the model’s reward and observation space distribution during training. These design combinations will be evaluated on continuous control tasks with established benchmarks.
Linear transformers for solving parametric partial differential equations
(2024) Hagnberger, Jan
The simulation of physical phenomena relies on solving Partial Differential Equations (PDEs), and Machine Learning models have increasingly addressed this task in recent years. PDEs often involve parameters influencing their evolution, prompting the development of models that consider these parameters as additional input. These parameter-conditioned models aim to generalize across different PDE parameters, replacing the need for multiple models trained on specific ones.
Transformer models have been achieving great success in Natural Language Processing (NLP), Speech Processing, and even in domains such as Computer Vision. Due to their ability to effectively model long-range dependencies in sequential data, their field of application is steadily increasing. Calculating attention via Scaled Dot-Product Attention in Vanilla Transformers is computationally expensive and scales quadratically with the input length. This leads to a bottleneck for very long sequences. To address this challenge, Linear Transformers have been introduced, substituting the Scaled Dot-Product Attention to achieve linear time and space complexity. Consequently, Linear Transformers have shown promising potential for processing very long sequences efficiently.
We investigate two approaches of utilizing Linear Transformers for solving PDEs and their associated problems. Moreover, we conduct a comprehensive comparison between our proposed transformer-based models and state-of-the-art models for solving parametric PDEs. The evaluation criteria include accuracy for short and long rollouts, memory consumption, and inference times. The results demonstrate that our proposed models perform competitively with the current state-of-the-art models, providing an efficient solution for PDE solving.
Training robust and generalizable quantum models
(2024) Berberich, Julian; Fink, Daniel; Pranjić, Daniel; Tutschku, Christian; Holm, Christian
Laser cooling of barium monofluoride molecules using synthesized optical spectra
(2024) Rockenhäuser, Marian; Kogel, Felix; Garg, Tatsam; Morales-Ramírez, Sebastián A.; Langen, Tim
Higher BTEX aromatic yield from ethanol over desilicated H,Zn-[Al]ZSM-5 catalysts
(2024) Dittmann, Daniel; Ileri, Alime; Strassheim, Dennis; Dyballa, Michael
The amount of BTEX aromatics obtained from the conversion of ethanol (ETA) is increased by combining ZSM-5 catalysts having optimum acidity with desilication and zinc ion exchange. Zinc leads to preferred dehydrogenation instead of hydrogen transfer. It decreases the share of paraffin products and increases BTEX contents (up to SBTEX = 50%) at the cost of lifetime. The latter can be increased via desilication. An ethylene feed increases lifetime and BTEX production as result of oxygenate absence. Combination of improvements resulted in a C2 conversion capacity of 206 g g-1 and a total yield of BTEX aromatics of 31.6 g g-1, which is about a factor of 2-3 times better than the respective values found for microporous, mesoporous, or microporous Zn-exchanged materials. In situ UV/vis spectra reveal that desilicated samples coke significantly slower than microporous samples, whereas Zn exchange supports the formation of coke. Thus, by a clever combination of suitable post-modifications, a significantly higher BTEX production from the primary source ethanol can be achieved.
Leveraging N-exo substituents to tune the donor/acceptor properties of mesoionic imines (MIIs)
(2024) Rudolf, Richard; Todorovski, Andrej; Schubert, Hartmut; Sarkar, Biprajit
In this work, we show two synthetic routes to substitute the Nexo position of mesoionic imines (MIIs). By Buchwald-Hartwig amination, 5-amino-1,2,3-triazoles can be arylated at the said position, showing the versatility of amino-triazoles as building blocks for MIIs. The reaction of MIIs with electrophiles (MeI, fluoro-arenes) highlights the nucleophilic nature of MIIs as even at room temperature aromatic C-F bonds can be activated with MIIs. By combining experimental methods such as Tolman/Huynh-electronic-parameter and crystallographic interpretations with theoretical calculations, we establish that MIIs expand the nucleophilicity scale of N-donors. Contrary to the flanking substituents on the triazole scaffold, the Nexo substituent heavily influences the donating ability of MIIs: electron-withdrawing substituents will dramatically decrease the donor strength of the MII ligand. We have now established ways to functionalise not only the triazole backbone but also the Nexo position. More importantly, we show here how the substitution pattern influences the electronic structure of MIIs. Such electronic tunability should make MIIs suitable for use in various fields of chemistry.