OPUS Online Publikationen der Universität Stuttgart
Not a member yet
17148 research outputs found
Sort by
Design of convex-shaped transparent heat spreaders for symmetrical cooling of a thin-disk laser crystal by mechanical pressing
No full textWe present the design for a double-sided cooling of a thin-disk laser crystal using transparent, convex-shaped heat spreaders. The thermal contact between laser crystal and heat spreaders is established without bonding by solely applying an axial compressive force. A model of the contact behavior taking into account the deformation of the surfaces of the heat spreaders due to bending under mechanical loading is presented. The scope and accuracy of the modeling is verified by numerical simulations using the finite-element method. Based on the modeling, an exemplary design of the heat spreader which has been used for the first experimental demonstration of this concept is carried out. Based on the model, an almost ideal scalability of the heat transfer coefficient is predicted.Projekt DEALSeventh Framework ProgrammeUniversität Stuttgar
The aluminum standard : using generative Artificial Intelligence tools to synthesize and annotate non-structured patient data
No full textBackground. Medical narratives are fundamental to the correct identification of a patient’s health condition. This is not only because it describes the patient’s situation. It also contains relevant information about the patient’s context and health state evolution. Narratives are usually vague and cannot be categorized easily. On the other hand, once the patient’s situation is correctly identified based on a narrative, it is then possible to map the patient’s situation into precise classification schemas and ontologies that are machine-readable. To this end, language models can be trained to read and extract elements from these narratives. However, the main problem is the lack of data for model identification and model training in languages other than English. First, gold standard annotations are usually not available due to the high level of data protection for patient data. Second, gold standard annotations (if available) are difficult to access. Alternative available data, like MIMIC (Sci Data 3:1, 2016) is written in English and for specific patient conditions like intensive care. Thus, when model training is required for other types of patients, like oncology (and not intensive care), this could lead to bias. To facilitate clinical narrative model training, a method for creating high-quality synthetic narratives is needed.
Method. We devised workflows based on generative AI methods to synthesize narratives in the German language to avoid the disclosure of patient’s health data. Since we required highly realistic narratives, we generated prompts, written with high-quality medical terminology, asking for clinical narratives containing both a main and co-disease. The frequency of distribution of both the main and co-disease was extracted from the hospital’s structured data, such that the synthetic narratives reflect the disease distribution among the patient’s cohort. In order to validate the quality of the synthetic narratives, we annotated them to train a Named Entity Recognition (NER) algorithm. According to our assumptions, the validation of this system implies that the synthesized data used for its training are of acceptable quality.
Result. We report precision, recall and F1 score for the NER model while also considering metrics that take into account both exact and partial entity matches. Trained models are cautious, with a precision up to 0.8 for Entity Type match metric and a F1 score of 0.3.
Conclusion. Despite its inherent limitations, this technology has the potential to allow data interoperability by using encoded diseases across languages and regions without compromising data safety. Additionally, it facilitates the synthesis of unstructured patient data. In this way, the identification and training of models can be accelerated. We believe that this method may be able to generate discharge letters for any combination of main and co-diseases, which will significantly reduce the amount of time spent writing these letters by healthcare professionals.Ministry for Economics, Labor and Touris
Int-HRL : towards intention-based hierarchical reinforcement learning
No full textWhile deep reinforcement learning (RL) agents outperform humans on an increasing number of tasks, training them requires data equivalent to decades of human gameplay. Recent hierarchical RL methods have increased sample efficiency by incorporating information inherent to the structure of the decision problem but at the cost of having to discover or use human-annotated sub-goals that guide the learning process. We show that intentions of human players, i.e. the precursor of goal-oriented decisions, can be robustly predicted from eye gaze even for the long-horizon sparse rewards task of Montezuma’s Revenge-one of the most challenging RL tasks in the Atari2600 game suite. We propose Int-HRL : Hierarchical RL with intention-based sub-goals that are inferred from human eye gaze. Our novel sub-goal extraction pipeline is fully automatic and replaces the need for manual sub-goal annotation by human experts. Our evaluations show that replacing hand-crafted sub-goals with automatically extracted intentions leads to an HRL agent that is significantly more sample efficient than previous methods.Projekt DEALDeutsche ForschungsgemeinschaftSchweizerischer Nationalfonds zur Förderung der Wissenschaftlichen ForschungBundesministerium für Bildung und ForschungEuropean Research Counc
Optimising light electric vehicle capabilities through a new battery main switch topology with emphasis on battery heating
No full textInvestigations of the use and optimization of LDS-MID technology in millimeter wave applications
No full textNeuere Hochfrequenzsysteme arbeiten bei immer höheren Frequenzen, was zum einen die Aufbautechnik schrumpfen, auf der anderen Seite die Verluste solcher Systeme wachsen lässt. Solche Systeme profitieren von 3D-Aufbautechniken, wie sie bei MID (Mechatronic Integrated Devices) und im speziellen LDS-MID (Laser Directstructured Mechatronic Integrated Devices) eingesetzt werden. Während sich bisherige Arbeiten, allen voran die Dissertationen von Frau Aline Friedrich und Frau Li Wang, mit der grundsätzlichen Machbarkeit eines solchen Systems beschäftigten, beinhaltet diese Arbeit eine Betrachtung aller relevanten technologischen Teilaspekte, um einen Einsatz von LDS-MID bis zu Frequenzen von 80 GHz zu ermöglichen. Als Erstes werden hierfür die Anforderungen an LDS-MID in mm-Wellenanwendungen zusammengefasst und eine große Anzahl von LDS-MID-Materialien hinsichtlich der relevanten Materialeigenschaften hin untersucht. Aus diesen Untersuchungen stellte sich heraus, dass das Material TECACOMP PEEK LDS black 1047045 das aktuell geeignetste Material zur Realisierung von mm-Wellen-LDS-MID ist. Basierend auf diesem Material wird die Anpassung spezifischer Eigenschaften durch einen partiellen Austausch des Matrixmaterials durch PEI sowie die dielektrischen Eigenschaften von PEEK LDS im mm-Wellenbereich untersucht. Durch Untersuchungen zum Spritzprägen von LDS-MID wird gezeigt, dass mit dieser Methode eine Reihe von relevanten Mikrostrukturen und großflächige Wandstärkenänderungen abgemustert werden können. Im darauffolgenden Kapitel schließt sich eine Detailcharakterisierung und Optimierung der Metallisierung auf PEEK LDS an. Da die meisten MMIC (Monolithic Microwave Integrated Circuit) im mmWellenbereich heutzutage in BGA-Packages ausgeliefert werden, folgt die Betrachtung der Lötung von BGA-Packages (Ball Grid Array Packages) auf LDS-MID. Dazu werden Entflechtungsstudien unterschiedlicher BGA-Pitches und ein Prozess zur lokalen Applikation von Lötstopplack vorgestellt. Dieser Aspekt wird dann durch die Beschreibung des Lötprozesses für BGA-Packages bis zu einem minimalen Pitch von 0,4 mm abgeschlossen. Die Validierung der LDS-MID Technologie für den mm-Wellenbereich erfolgt dann durch die Charakterisierung von Übertragungsleitungen auf spritzgeprägten LDS-MID bis 30 GHz sowie auf folienextrudiertem PEEK LDS im Frequenzband zwischen 60 und 90 GHz. Abgeschlossen wird die Arbeit dann mit der Beschreibung eines Radardemonstrators mit einer Arbeitsfrequenz bei 80 GH
Next-generation sustainable composites with flax fibre and biobased vitrimer epoxy polymer matrix
No full textThis work presents the development of two vanillin-based vitrimer epoxy flax fibre-reinforced composites, with both the VER1-1-FFRC (a vitrimer-to-epoxy ratio of 1:1) and VER1-2-FFRC (a vitrimer-to-epoxy ratio of 1:2), via a vacuum-assisted resin infusion. The thermal and mechanical properties of the resulting vitrimer epoxy flax composites were characterised using thermal gravimetric analysis (TGA), differential scanning calorimetry (DSC), dynamic mechanical analysis (DMA), and mechanical four-point bending tests, alongside studies of solvent resistance and chemical recyclability. Both the VER1-1-FFRC (degradation temperature Tdeg of 377.0 °C) and VER1-2-FFRC (Tdeg of 395.9 °C) exhibited relatively high thermal stability, which is comparable to the reference ER-FFRC (Tdeg of 396.7 °C). The VER1-1-FFRC, VER1-2-FFRC, and ER-FFRC demonstrated glass transition temperatures Tg of 54.1 °C, 68.8 °C, and 83.4 °C, respectively. The low Tg of the vitrimer composite is due to the low crosslink density in the vitrimer epoxy resin. Particularly, the crosslinked density of the VER1-1-FFRC was measured to be 319.5 mol·m−3, which is lower than that obtained from the VER1-2-FFRC (434.7 mol·m-3) and ER-FFRC (442.9 mol·m-3). Furthermore, the mechanical properties of these composites are also affected by the low crosslink density. Indeed, the flexural strength of the VER1-1-FFRC was found to be 76.7 MPa, which was significantly lower than the VER1-2-FFRC (116.2 MPa) and the ER-FFRC (138.3 MPa). Despite their lower thermal and mechanical performance, these vitrimer composites offer promising recyclability and contribute to advancing sustainable composite materials.Swinburne University of TechnologyAustralian Department of Industry, Innovation and Scienc
Unveiling the origin of the yield stress anomaly in L12 intermetallics via atomistic approaches
No full textSampling music spaces with generative AI
No full textRecent advances in deep learning have led to powerful models for symbolic music generation, but their creative potential often comes at the expense of user controllability. To provide a more controllable environment for AI-assisted composing, we present a visual approach based on latent embeddings from models, creating an interactive two-dimensional music space in which users can generate melodies in desired areas. By applying dimensionality reduction to embeddings from state-of-the-art symbolic music generation models, melodies are mapped into a scatterplot where proximity reflects similarity. We develop an interactive framework where users can generate, select, and explore melodies directly in the music space. Within this framework, we implement several generation techniques based on the MusicVAE, Pop Music Transformer (REMI), and FIGARO models, and evaluate their effectiveness in filling user-defined regions of the space. Quantitative results show that MusicVAE’s similar and interpolate methods most reliably generate samples close to the targeted area, while REMI and FIGARO produce greater diversity at the cost of precision. A qualitative analysis further highlights how dimensionality reduction methods, parameter settings, and spatial density influence the outcome of generation. Results show that techniques based on the MusicVAE model generate melodies that are both in close proximity and musically similar to their surroundings, making it the most applicable method to generate samples in desired areas of the music space. Our work contributes to the development of visual, interactive methods for human-AI co-creativity in music, emphasizing controllability, exploration, and inspiration in the composition process.Die jüngsten Fortschritte im Bereich des Deep Learning brachten leistungsstarke Modelle für die symbolische Musikgenerierung hervor, ihr Potenzial geht jedoch häufig zulasten der Beinflussbarkeit durch den Nutzer. Um diesem Problem zu entgegnen, stellen wir einen visuellen Ansatz für KI-unterstütztes Komponieren vor, der mithilfe latenter Embeddings generativer Modelle einen interaktiven zweidimensionalen Musikraum erzeugt, in welchem Nutzer neue Melodien in ausgewählten Bereichen erzeugen können. Durch Dimensionalitätsreduktion werden die hochdimensionalen Embeddings auf einem Scatterplot abgebildet, bei welchem die Nähe zweier Punkte deren musikalische Ähnlichkeit widerspiegelt. Auf dieser Grundlage entwickeln wir ein Framework, in dem Nutzer Melodien direkt im Musikraum generieren, auswählen und erkunden können. Im Rahmen unserer Arbeit werden mehrere Generationsverfahren implementiert, die auf den Modellen MusicVAE, Pop Music Transformer (REMI) und FIGARO basieren. Anschließend evaluieren wir deren Effektivität zur gezielten Befüllung benutzerdefinierter Bereiche des musikalischen Raums. Die quantitative Ergebnisse zeigen, dass MusicVAEs similar- und interpolate-Methoden am zuverlässigsten Ausgaben erzeugen, die nahe am Zielbereich liegen, während REMI und FIGARO größere Vielfalt liefern, die auf die Kosten der Präzision gehen. Eine qualitative Analyse verdeutlicht zudem, wie Dimensionalitätsreduktion, Parametereinstellungen und räumliche Dichte den Generationsprozess beeinflussen. Die Ergebnisse zeigen, dass die Verfahren basierend auf dem MusicVAE-Modell Melodien generieren, die sowohl nah als auch musikalisch ähnlich zu ihren umliegenden Punkten sind. Daraus erschließt sich, dass sich dieses Modell zur Generierung in gezielten Bereichen am besten eignet. Mit dieser Arbeit leisten wir einen Beitrag zur Weiterentwicklung visueller und interaktiver Methoden für die Mensch-KI-Co-Kreativität in der Musik und legen den Fokus dabei insbesondere auf Steuerbarkeit, Exploration und Inspiration im Kompositionsprozess