Eldorado - Repository of the TU Dortmund
Resources for and from Research, Teaching and Studying
This is the institutional repository of the TU Dortmund. Ressources for Research, Study and Teaching are archived and made publicly available.
Communities in Eldorado
Select a community to browse its collections.
Recent Submissions
The effect of quantum confinement on the spin properties of lead halide perovskites probed by resonant Raman spectroscopy
(2025) Harkort, Carolin Sophie; Yakovlev, Dmitri; Reiter, Doris
Lead halide perovskites have emerged as exceptional semiconductor materials for photovoltaic and optoelectronic applications, offering easy tunability and lower production costs than conventional semiconductors. Their band gap energy can be adjusted through compositional changes, particularly by modifying the halide content, and through quantum confinement in low-dimensional systems. While the effects of com- position and dimensionality on optical properties are well established, their influence on spin properties is far from being well understood. In this work, the technique of spin-flip Raman spectroscopy is employed to investigate three-, two-, and zero-dimensional lead halide perovskites, focusing on a key band structure parameter defining the coupling of spins to external magnetic fields: the Landé 𝑔-factor. In particular, the impact of quantum confinement on the carrier 𝑔-factor is examined in Ruddlesdon-Popper type two- dimensional perovskites and zero-dimensional CsPbBr3 perovskite nanocrystals. The dependence of their 𝑔-factors on the effective band gap energy is compared to the universal dependence of the electron and hole 𝑔-factors in three-dimensional lead halide perovskites. This work reveals that while the general trend of both electron and hole 𝑔-factors follows the bulk dependence, significant deviations in their absolute values occur in two- and zero-dimensional lead halide perovskites, highlighting the pronounced impact of quantum confinement on spin properties. From a technological perspective, it is particularly interesting that the 𝑔-factor can be engineered by adjusting the number of inorganic layers in 2D perovskites or the size of the nanocrystals. Spin-flip Raman spectroscopy also reveals the domain structure of a bulk MAPbI3 single crystal by identifying the presence of domains with different crystal orientations through the 𝑔-factor anisotropy. Furthermore, rare double spin-flip processes involving two electrons or two holes are detected. Next to spin-flip processes, confined acoustic phonon modes are discovered in the Raman spectra of CsPbBr3 nanocrystals. A comparison of experimental results and density functional theory calculations enable the identification of these phonon modes and offers a complementary optical tool to probe structural properties such as the shape, structural phase, and size of the nanocrystals that are, in turn, key to understanding spin interactions.
Compilation-based explainability of tree ensembles
(2025) Murtovi, Alnis; Steffen, Bernhard; Jansen, Nils
Machine learning, particularly deep learning, has achieved remarkable success in areas such as image recognition, natural language processing, and speech recognition.
However, for structured or tabular data, tree ensemble approaches, such as random forests and gradient boosted trees, often outperform deep learning-based approaches.
Although they perform strongly, tree ensembles are in general considered to be black boxes, because the complexity of combining multiple decision trees makes it difficult to trace the reasoning behind individual predictions.
Recent advancements in Explainable Artificial Intelligence have presented heuristic post-hoc explanation methods like LIME and SHAP.
Nevertheless, these methods often rely on a model's input-output behavior and therefore only approximate how it internally arrives at its predictions.
As a result, there is an urgent need for efficient and precise approaches to explaining ensembles, especially in domains where safety or fairness is critical.
To tackle the interpretability gap, this thesis explores compilation-based techniques that transform an entire tree ensemble into a single, semantically equivalent structure such as a directed acyclic graph.
This single graph representation reveals the ensemble's underlying logic, making it amenable to formal analysis.
Once compiled, the model's internal logic becomes more transparent, allowing efficient generation of formal explanations, as well as support for verification tasks such as pre- and postcondition checks and model equivalence checking.
One significant advantage is that, after the one-time cost of building the unified representation, subsequent explanations can be generated efficiently.
This makes the proposed solutions particularly well-suited for real-time or interactive settings where many explanations are requested in sequence.
The main focus of this thesis is efficiency and scalability.
Existing compilation-based approaches can be very expensive on large ensembles.
To address this, the thesis introduces novel compilation algorithms and optimizations, significantly reducing transformation time and memory usage while maintaining exact equivalence with the original tree ensemble.
The experimental results show over an order of magnitude speedup in model compilation and multiple orders of magnitude in explanation generation compared to state-of-the-art solver-based approaches.
Furthermore, the thesis introduces a user-friendly, web-based tool (Forest GUMP) that allows non-experts to train, visualize, verify, and explain tree ensembles interactively.
Overall, these contributions advance the field of explainable AI by delivering efficient, formally grounded, and practical solutions for tree ensemble interpretability and explainability.
Realistic virtual humans for VR therapy of body image disorders
(2025) Wenninger, Stephan; Botsch, Mario; Stamminger, Marc
Stephan Wenninger - Realistic Virtual Humans for VR Therapy of Body Image Disorders
Abstract:
This thesis presents methods for reconstructing and modifying realistic personalized virtual humans to be employed in the context of a VR-based body image disorder therapy system.
We start by presenting a method for generating virtual humans from monocular smartphone cameras, thereby lowering the hardware requirements and increasing the availability of personalized virtual humans compared to other methods, which typically depend on elaborate photogrammetry rigs.
In a user study, we investigate the perception of the resulting virtual humans by scanning people with both the low-cost smartphone-based method and a standard multi-view stereo photogrammetry rig.
Participants then embody and rate both virtual humans in a virtual mirror exposure scenario.
The results show, that both virtual humans are perceived similarly, indicating that our smartphone-based method presents a viable alternative to expensive photogrammetry rigs.
For employing realistic virtual humans in body image therapy, we present a method for modifiying the body weight of the virtual humans in real-time.
Users of the VR-based therapy system then embody a personalized avatar in a virtual mirror exposure scenario and are given active control over the avatar's body shape, enabling researchers to investigate the potential of VR-based therapy and gain insight into possibly occurring body image disorders.
To improve on the purely surface-based body weight modification model, the second part of this thesis focuses on anatomical representations of virtual humans.
We present a method for inferring anatomical details from a given skin surface in less than a minute.
To this end, we derive a three-layered anatomical model, consisting of a skin, muscle, and skeleton layer, from a commercial high-resolution anatomical model.
We then learn a model for predicting body composition, i.e., fat and muscle mass, from a given skin surface and fit the template model to a large database of surface scans while conforming to the estimated body composition.
The original high-resolution anatomical structures are transferred to the resulting fit via a triharmonic space warp.
Finally, we use the inferred anatomical data to learn an anatomically constrained volumetric human shape model.
We enlarge our training data to the full Cartesian product of all skeleton shapes and all soft tissue distributions using physically plausible volumetric deformation transfer.
A self-supervised learning technique then produces two separate latent parameter sets, allowing us to sample different soft tissue distributions over the same skeleton shape and vice versa.
The resulting anatomical model additionally facilitates fast skeleton inference and semantic localized shape modification.
Erwerbsverläufe, Gesundheit und der Altersübergang von Erwerbstätigen in der Einfacharbeit
(2025) Kaboth, Arthur; Hünefeld, Lena; Himmelreicher, Ralf
Die kumulative Dissertation befasst sich mit den Erwerbsverläufen, der Gesundheit, der Arbeitsfähigkeit sowie dem Übergang in die gesetzliche Altersrente von Beschäftigten in der sogenannten Einfacharbeit, Tätigkeiten, die keinen schulischen oder beruflichen Ausbildungsabschluss erfordern.
Die Erweiterung des aktuellen Forschungsstands von querschnittlichen zu längsschnittlichen Analysen können zum einen das negative Bild der Einfacharbeit bestätigen: Einfacharbeit ist geprägt von Erwerbsunterbrechungen, gesundheitlichen Risiken, niedriger Arbeitsfähigkeit sowie vorzeitigen Renteneintritten. Zum anderen zeigen die Ergebnisse, dass ein Anteil der Beschäftigten in der Einfacharbeit in zeitlich stabiler Vollzeittätigkeit (Normalarbeitsverhältnis) tätig ist, Übergänge in die Facharbeit ermöglicht werden, Renteneintritte für einen Anteil später erfolgen als in der Facharbeit sowie psychische Belastungsreduzierung ein Motiv für eine Beschäftigung in Einfacharbeit darstellen kann. Insgesamt bedarf es allerdings Maßnahmen und Initiativen zur Verbesserung der Arbeits- und Gesundheitssituation sowie weiterer Forschung im Forschungsfeld Einfacharbeit.
„Die Treppe wird immer von oben nach unten gekehrt“ (?)
(2024) Wolf, Lisa Marie; Kuhl, Jan; Moser, Vera
Mit der Ratifizierung der UN-BRK sind die deutschen Bundesländer dazu aufgefordert, ein inklusives Schulsystem zu entwickeln. Eine zentrale Steuerungshandlung ist dabei der Einsatz sonderpädagogischer Lehrkräfte an Regelschulen. Für die Lehrkräfte dieser Profession, die traditionell eng an die Institution Förderschule gekoppelt ist, bedeutet dies nicht nur eine Veränderung ihres Arbeitsortes, sondern auch eine Veränderung beruflicher Anforderungen und Aufgaben. Die Dissertation untersucht, wie und auf welchen Ebenen des Schulsystems der Einsatz sonderpädagogischer Lehrkräfte an inklusiven Grundschulen gesteuert wird und welche Konsequenzen sich daraus für die sonderpädagogischen Lehrkräfte ergeben. Der Vergleich der Bundesländer Nordrhein-Westfalen und Hessen ermöglicht dabei Einblicke in verschiedene Strukturen und Steuerungsmodelle. Grundlage der Analyse bilden Interviews mit insgesamt 21 sonderpädagogischen Lehrkräften, die durch weitere Datenquellen (offizielle Dokumente zur Steuerung, quantitative Befragungen von Grundschulleitungen und sonderpädagogischen Lehrkräften) ergänzt werden. Die Daten stammen aus dem BMBF-geförderten Projekt FoLis (FKZ: 01NV1718A/B, 2018–2021). Die Ergebnisse zeigen, dass in der Steuerung der Bundesländer Unterschiede bestehen, aus denen sich teils bundeslandspezifische Herausforderungen für die sonderpädagogischen Lehrkräfte sowie Ansatzpunkte für Nachsteuerungen ergeben. Deutlich wird insgesamt aber auch die Bedeutung schulinterner Rahmenbedingungen sowie der interdisziplinären Kooperation.