Recent Submissions
Mapping the nonribosomal specificity code through promiscuity-guided A-domain engineering
(2026) Stanišić, Aleksa; Svensson, Carl-Magnus; Müll, Maximilian; Bernal, Freddy A.; Zeihe, Hannah; Ettelt, Ulrich; Kries, Hajo
Nonribosomal peptide synthetases (NRPSs) assemble bioactive peptides from various building blocks. The binding pocket residues governing building block specificity have allowed prediction of NRPS products but not design of specificity. A reason for this failure has been ignorance of NRPS multispecificity. Here, we employ a hydroxamate assay (HAMA) to determine multispecificity for mutant libraries of the adenylation (A-)domain in module SrfAC of surfactin synthetase. A multispecific version of SrfAC is developed and its functional flexibility probed by fully randomizing 15 residues around the active site. We identify mutations with profound impact on specificity revealing remarkable evolvability and explain the effect of a selected mutant by computational modelling. Statistical analysis of the specificity divergence caused by 285 point mutations has revealed an outstanding influence of three sequence positions on specificity, which provides a roadmap for NRPS engineering. Our results suggest that promiscuity drives neofunctionalization of A-domains and mimicking this process will help to design valuable peptides in the lab.
The anionically all‐boron‐based anti‐perovskites Cs3[BH4][B12H12] and Cs3[BF4][B12H12]
(2025) Tiritiris, Ioannis; Hakkak, Rouzbeh Aghaei; Ranjbar, Maryam; Schleid, Thomas
The anionically all‐boron‐based colorless anti‐perovskites Cs3[BH4][B12H12] and Cs3[BF4][B12H12] can be obtained by isothermal evaporation from aqueous solutions of the underlying pseudo‐binary components Cs2[B12H12] and Cs[BH4] or Cs[BF4]. At ambient conditions, Cs3[BH4][B12H12] crystallizes with the lattice parameters a = 1054.89(8) and c = 1197.12(9) pm (Z = 3) in the trigonal space group R3¯m. According to differential scanning calorimetry studies, it transforms reversibly into a cubic structure (a = 744.50(6) pm, space group: Pm3¯m, Z = 1) above 310°C. Both forms of Cs3[BH4][B12H12] represent derivatives of the cubic anti‐perovskite arrangement, while the orthorhombic crystal structure of Cs3[BF4][B12H12] (a = 874.69(5), b = 1052.71(6) and c = 1724.52(9) pm, space group: Pnma, Z = 4) adopts the hexagonal anti‐perovskite motif. So quasi‐icosahedral [B12H12]2- anions and Cs+ cations assemble to a mixed cubic or hexagonal closest packed framework, respectively, in which the quasi‐tetrahedral [BH4]- or [BF4]- anions (A) occupy the quasi‐octahedral voids erected only by cations in a way that ∞3{[ACs6/2]2+} or ∞1{[ACs6/2]2+} substructures of all‐vertex connected [ACs6]5+ octahedra occur in the first, but trans‐face sharing ones in the second case.
Translating property graphs to RDF-star using graph generating dependencies
(2025) Ganguly, Samhita
The Property Graph (PG) [Ang18] model is widely used in modern data systems for its intuitive handling of nodes, relationships, and associated properties. However, PG lacks formal semantics and interoperability, particularly with web-native frameworks such as RDF and RDF-star [ABPS+12; KFH22; NYT+19]. While existing systems like Neosemantics [RCK22] and mapping languages such as RML and RML-star [BCML; Con+12b; DAI+21] allow PG-to-RDF transformation, they either rely on user-defined mappings or impose rigid ontology-based encodings [DVC+14; RCK22; TAT20]. To address these limitations, this thesis proposes the Graph Generating Transformer (GGT): a modular, rule-based transformation system that converts PG data into RDF-star using a fixed, reusable set of declarative transformation rules.
The core innovation lies in applying Graph Generating Dependencies (GGDs) to define transformation semantics that preserve both the structure and metadata of PGs. Unlike prior approaches, GGT does not require users to define mappings per dataset and avoids schema reification. It supports RDF-star features such as quoted triples for representing edge properties and ensures datatype-annotated output for semantic querying.
The system is implemented in Python with an extensible architecture comprising loaders, rule applicators, and RDF-star generators. Evaluation on WatDiv benchmark datasets and a custom evaluation set demonstrates that GGT scales to thousands of triples while maintaining clear semantic output. Runtime and memory consumption are analyzed across data sizes to assess practical viability. This thesis demonstrates that a GGD-based rule set can serve as a robust foundation for interoperable graph transformation to RDF-star, bridging the gap between PG and RDF-Star semantics [SYF24a].
Intelligent view sampling for outdoor scalability
(2026) Renganathan, Harish
Photorealistic 3D reconstruction with neural rendering methods depends heavily on the quality of the captured image data. How to collect that data well has received far less attention than the reconstruction algorithms themselves. IntelliCap recently tackled this by guiding smartphone users during capture in real time, overlaying striped patterns over unscanned surfaces and semi-transparent spheres over objects that are difficult to reconstruct. The system works, but all the heavy computation runs on a remote server, introducing a per-detection-cycle round-trip delay that makes real-time guidance impractical, and the system stops working entirely without a reliable connection.
This thesis describes a redesign of IntelliCap's capture guidance for entirely on-device execution. YOLOv8s replaces Detectron2 for object detection, running under a seven-frame skip strategy that amortises inference cost across rendered frames to stay within the real-time budget. A pre-computed per-class difficulty table replaces the server-side language model used by IntelliCap for object importance scoring, eliminating that inference cost entirely. A two-tier depth hierarchy handles the full range of mobile depth failure modes without placing spatially misleading spheres: ARCore's Depth API is tried first, then raymarching into an accumulated TSDF volumetric map, with sphere creation skipped when both sources fail. Coverage-aware stripe rendering samples the TSDF weight atlas in a GPU shader, giving the overlays a persistent spatial memory that doesn't flicker with depth dropouts or fast camera movement.
The fully on-device design eliminates round-trip network latency entirely, enabling operation without any server connection. No network connection is needed. A four-participant user study returned strong scores on sphere 3D accuracy and stability, with more moderate results on coverage speed and quality. Three participants cited the fully mobile execution as the system's standout feature. Common improvement requests centred on sphere clutter in dense scenes and the absence of any global navigation cue for prioritising which unscanned regions to visit next.
Intelligent view sampling with an AR headset for novel view synthesis
(2026) Dhulappa Itagi, Sampathkumar
Recent advances in neural rendering, particularly 3D Gaussian Splatting, have enabled real-time novel view synthesis suitable for immersive applications. However, acquiring input images of sufficient quality remains a bottleneck, as accurate reconstruction demands both broad spatial coverage and diverse viewing angles-requirements that are not intuitive for human operators. Prior work, notably IntelliCap, has demonstrated that augmented reality guidance combining spatial coverage feedback with object-aware angular sampling can improve capture quality on smartphone platforms. However, smartphones impose constraints: a narrow field of view, single-handed operation, and the absence of stereoscopic depth cues, all of which limit spatial awareness during capture.
This thesis investigates whether migrating such a guidance system to a head-mounted stereoscopic display-the Meta Quest 3-can address these limitations. The transition introduces distinct design challenges, including rendering guidance overlays consistently across both eye views, adapting interaction from phone-pointing to head-oriented scanning, and executing the full pipeline entirely on-device without server offloading. The approach is evaluated through a controlled reconstruction analysis and an exploratory expert review with four domain experts in augmented reality.
The combined spatial and angular guidance condition yields consistent improvements in reconstruction quality over both unguided and spatial-only capture, with experts confirming that stereoscopic wide-field-of-view presentation enhances spatial awareness during scanning. While the small expert pool and closed-vocabulary object detection represent current limitations, the results demonstrate the feasibility and potential of head-mounted AR guidance for neural rendering capture tasks.
Untersuchung und Optimierung einer Mikrogasturbine für die Anwendung als automobiler Primärenergiewandler
(Stuttgart : Deutsches Zentrum für Luft- und Raumfahrt, Institut für Verbrennungstechnik, 2026) Kislat, Oliver; Aigner, Manfred (Prof. Dr.-Ing.)
Die Elektrifizierung des Automobilsektors wird als wichtiger Schritt zur Erreichung der Klimaneutralität im Transportsektor betrachtet. Trotz hohem Entwicklungstempo gibt es Vorbehalte bei den erzielbaren Reichweiten und Ladezeiten. Für Fahrzeuge mit hohen Reichweitenanforderungen können Hybridantriebe eine Vielzahl von Möglichkeiten eröffnen. So bietet der Einsatz von elektrischen Antrieben die Möglichkeit einen Primärenergiewandler entkoppelt von den erforderlichen Dynamiken und Spitzenleistungen in einem günstigen Betriebspunkt zu betreiben. Das ermöglicht den Einsatz alternativer Technologien, wie Brennstoffzellen oder Mikrogasturbinen.
Mikrogasturbinen sind in der dezentralen Energieversorgung, insbesondere in Bereichen wo Kraftstoffflexibilität oder die Auskopplung von Prozesswärme erforderlich sind, bereits etabliert. Ihr vergleichsweise geringer elektrischer Wirkungsgrad nach aktuellem Stand der Technik und hohe Anschaffungskosten stehen dem Einsatz im Automobilbereich entgegen. Die Vorzüge von Mikrogasturbinen liegen zur Zeit in potenziell geringen Unterhaltskosten und niedrigen Schadstoffemissionen. Aufgrund ihres modularen Aufbaus bieten sie viele Ansätze zur Komponentenoptimierung für eine spezifische Anwendung.
In dieser Arbeit wurde auf Basis einer kommerziell verfügbaren Mikrogasturbine vom Typ Capstone® C30 mit 30 kW elektrischer Leistung bei einem elektrischen Wirkungsgrad von 25 % die Eignung und das Optimierungspotenzial von Mikrogasturbinen für die Anwendung als automobiler Primärenergiewandler experimentell und numerisch untersucht. Dazu wurde ein Demonstrator-Prüfstand aufgebaut und mit detaillierter Instrumentierung an den wichtigen Kreislaufkomponenten, einer Abgasanalyse und einer Verdichterzapfluft ausgestattet. Anhand der so instrumentierten C30 wurde das Gesamtsystem, sowie die einzelnen Komponenten stationär und transient über den gesamten Betriebsbereich vermessen. Der experimentelle Datensatz wurde dazu genutzt, ein im institutseigenen Kreislaufsimulationstool MGTS3 bestehendes Modell der C30 zu validieren. Mit diesem wurden dann Parameterstudien zu Kenngrößen wie Wirkungsgrad oder Druckverlust der Systemkomponenten durchgeführt. Die Ergebnisse wurden hinsichtlich ihrer Wirksamkeit und Umsetzbarkeit zur Optimierung der Mikrogasturbine bewertet. Es wurde gezeigt, dass durch die Umsetzung einzelner Maßnahmen mit der C30 elektrische Wirkungsgrade von 30 % erreicht werden können.
Als Optimierungsmaßnahmen wurden im Rahmen dieser Arbeit der Rekuperator, die Brennkammer und deren strömungstechnische Integration identifiziert. Deshalb wurde ein neuer Rekuperator konzipiert und durch einen Zulieferer ausgelegt und additiv gefertigt. Es wurden Konzepte zur Integration eines FLOX®-basierten Brennkammersystems aufgezeigt und die konstruktive Umsetzung zur schrittweisen Integration von Rekuperator und Brennkammer beschrieben. Zur Untersuchung der Fahrzeugintegration der Mikrogasturbine wurden eine Zuluftkonditionierung und Abgasblenden ausgelegt. Die Gasturbine wurde mit dem neuem Rekuperator in Betrieb genommen und über den gesamten Drehzahlbereich charakterisiert. Es zeigte sich jedoch, dass die modifizierte Gasturbine interne Leckagen aufweist und die Druckverluste des gefertigten, neuentwickelten Rekuperators die Auslegungswerte deutlich übersteigen, sodass keine Verbesserung der Leistungskennwerte erzielt wurde. Die Vorkonditionierung und Abgasblenden kamen bei der modifizierten Mikrogasturbine zum Einsatz, um den Einfluss von Verdichtereintrittsbedingungen und Abgasdruckverlusten zu charakterisieren.
Um die experimentell ermittelten Kraftstoffverbräuche und Emissionen sowie deren Optimierungspotenziale in fahrzeugrelevanten Dimensionen einzuordnen, wurde in MATLAB® Simulink® ein Fahrzyklussimulator entwickelt. Neben den Bewegungsgleichungen wurden das stationäre und transiente Verhalten der C30 Gasturbine auf Basis der Messergebnisse modelliert. Ausgewertet wurden die Fahrzyklen nach dem WLTP und einem Langstreckentest basierend auf der gesamten Tank- und Batteriekapazität. Mit dem Simulator wurden verschiedene Betriebskonzepte der Mikrogasturbine in Abhängigkeit des Leistungsbedarfs und des Ladezustands der Batterie untersucht. Anhand von Parametervariationen globaler Gasturbinenkennwerte wurde der Einfluss der ermittelten Optimierungspotenziale auf das Gesamtfahrzeug übertragen.
So bietet diese Arbeit eine ganzheitliche Betrachtung des Einsatz- und Optimierungspotenzials von Mikrogasturbinen in PKW. Ausgehend von der experimentellen Erfassung des Ist-Zustands einer Mikrogasturbine in der relevanten Leistungsklasse wurden durch numerische Simulation Optimierungspotenziale identifiziert. Diese wurden in einem Demonstrator umgesetzt und experimentell charakterisiert. Die Erkenntnisse aus dem Optimierungsprozess wurden abschließend in einem Fahrzyklussimulator im Gesamtfahrzeug bewertet. Damit kann diese Arbeit als Grundlage für künftige Entwicklungsstrategien hinsichtlich des Einsatzes von Mikrogasturbinen in Kraftfahrzeugen dienen.
A framework for robust and explainable gear fault diagnosis
(2025) Schmid, Tobias
Gear fault detection is a sub field of predictive maintenance, that deals with automatically identifying gear faults, often using machine learning on vibration data. One of these gear faults is pitting damage, which deals with small pits on the gear tooth surface. Pitting detection models should ideally have three desirable qualities in a practical detection setting. First, they should be trained in an unsupervised manner. Next, they should be robust to intentional changes in gear behavior, like a change in gear speed or torque. Finally, their decisions should be interpretable to human decision makers. The objective of this thesis is to aid the development of such models. We attempt to achieve this objective by providing a theoretical framework that allows for the analysis of robustness and improves explainability in a lightweight manner. An analysis of 3 unsupervised models is provided as a demonstration of the capabilities of the framework. The evaluation of the models with the framewokr demonstrates the utility of Isolation Forests in pitting detection. It further showcases how simple input features like mean amplitude can be used in pitting detection. While no perfectly robust and explainable model is found, the evaluation clearly demonstrates the utility of the framework in aiding the evaluation of machine learning models with regards to robustness and explainability.
Advancing deep generative models for improved visual defect recognition in optimized production environments with limited data
(Stuttgart : Fraunhofer-Institut für Produktionstechnik und Automatisierung IPA, 2026) Wang, Ru-Yu; Huber, Marco (Univ.-Prof. Dr.-Ing. habil.)
The landscape of industrial manufacturing processes has seen remarkable changes since the first industrial revolution. The latest phase of this change is called the fourth industrial revolution and is characterized by interconnected and autonomous systems. It focuses on improving production efficiency and product quality through advanced information and communication technologies, allowing machines to self-correct and adapt to unforeseen circumstances.
A key challenge in this context is defect recognition, which is crucial for quality control. Traditional manual inspection methods are time-consuming and carry the risk of human error. As a result, Automatic Vision Inspection (AVI) systems have emerged as a promising alternative, providing consistent and measurable mechanisms for defect detection in real-time.
However, the success of such systems is heavily dependent on advanced software algorithms, especially those that utilize deep learning. These deep learning algorithms require large amounts of well-balanced training data, which are not available in many cases due to optimized production settings with few defective products.
In this thesis, the lack of data is tackled by leveraging advanced Deep Generative Models (DGMs) to create synthetic data of defective products. This synthetic data is then used to improve the performance of deep learning-based defect recognition algorithms. However, the generative models used to create synthetic data also need a lot of training data. This requirement presents a major challenge.
To address this challenge, two innovative DGMs are introduced in the thesis. The first model, Defect-Transfer GAN (DT-GAN), capitalizes on the similarities in defects among various products, effectively reducing the data requirements for individual products. The second model, named STyled LAYout Diffusion (STAY Diffusion), employs a flexible bounding box layout to generate defects at predetermined locations of a product. This feature ensures that defects are generated only where they can occur on a real product. Moreover, it allows for learning the pattern of a defect type across multiple products, which further decreases the need for extensive data.
In addition, the effectiveness of synthetic data generated by DGMs is investigated thoroughly in this thesis. The analysis focuses on the utility of this data for training deep learning-based systems like AVI systems. It provides insights into the discrepancies between synthetic and real-world data that persist despite the rapid advancements in DGMs. To counteract this synthetic-to-real gap, a new regularization technique is proposed. This technique combines knowledge from informative pretrained models with a few real samples to enhance performance in discriminative tasks, such as defect recognition.
All models and methods outlined in this thesis have undergone extensive evaluation on various academic and real-world industrial datasets. The findings demonstrate that the proposed DGMs produce highly realistic synthetic data, significantly improving the performance of deep learning-based models in solving defect recognition tasks. Moreover, applying the proposed regularization technique to the images generated by DGMs further enhances their effectiveness as training data. This leads to better performance of the deep learning-based recognition models without necessitating changes to the DGMs.
Scaling and optimizing crowdsourcing for high-accuracy geospatial data acquisition
(2026) Collmar, David; Sörgel, Uwe (Prof. Dr.-Ing.)
Optimised design and operating strategies for latent heat-thermal energy storage for steam generation
(2026) Dietz, Larissa; Vandersickel, Annelies (Prof. Dr.)
The characteristic high energy density within a narrow temperature interval makes latent heat-thermal energy storage (LH-TES) systems a promising technology in the field of carbon-neutral process-steam generation. State-of-the-art LH-TES systems are based on the tube-and-shell heat exchanger design with the storage material, the so-called phase change material (PCM), in the shell and the heat transfer fluid (HTF) flowing in the tubes. The challenge in storage operation lies in the low thermal conductivities of potential PCMs, which result in a characteristic transient power profile that strongly depends on the state of charge of the storage system. To meet the specific requirements of an application, suitable operating strategies are needed in which the heat transfer characteristics of the storage are combined with appropriate control of the steam flow. So far, the focus in the development of LH-TES for process-steam supply was on the heat transfer processes within the storage. This thesis provides a comprehensive experimental, analytical and numerical analysis of the interaction between the LH-TES and the processes in the steam flow. On this basis, it proposes an extended design method for the thermo-economic optimisation of steam-driven LH-TES systems by also considering partial-load operation. Analytical studies show the effects of the HTF mass flow rate and the transient thermal resistances on the storage power. A numerical storage model is used to determine strategies for controlled storage operation. These proposed strategies are assessed on a finned single-tube PCM steam generator test rig with PlusICE® A133 as the storage material, precise control of the HTF parameters and detailed monitoring of the storage tank. The results of the experiments under full and partial load with water/steam between 1.4 bar and 6.7 bar and mass flow rates between 1.26 kg/h and 4.30 kg/h at inlet temperatures between 20 °C and 200 °C demonstrate the feasibility of these operating strategies and show that pressure and mass flow rate on the fluid side can be used effectively to shape the power profile of LH-TES. Furthermore, the numerical storage design model is validated with experimental charging results in the operating range under consideration. Finally, this work shows that for applications where sufficiently powerful fins are not available or economically not feasible to provide stable steam pressures, temperatures and mass flow rates, partial-load strategies with reduced mass flow rates can be used to provide process-steam at increased material efficiency. It is a first step towards a more integrated approach to the development of LH-TES systems and complements the existing approach of heat transfer enhancement on the storage side in LH-TES under full-load operation.