Converging calculations is a common need in the ab initio materials-science community.
This tedious and resource-intensive process can be largely avoided if well-validated
recommendations are available. In order to create a recommender system to assist
users, benchmark data are required. This project addresses this need. It evaluates the
convergence behavior of electronic properties for a dataset of 10 materials that are
promising for optoelectronic applications.

Computational tools such as Molecular Dynamics (MD) have revolutionized the way we study biomolecules; however, they are severely limited by the computational cost of running simulations on biological time- and length-scales. Various coarse-grained (CG) models have been developed which rely on simpler representations of molecular systems than atomistic MD. While these models are difficult to configure using physical intuition, we have shown that by using state-of-the-art machine learning methods, it is possible to design accurate and efficient CG models which can correctly reproduce protein dynamics. By enhancing both our training dataset and network architecture, we hope to produce a “universal” CG model to study biological systems.

Within Text+, the NFDI consortium dedicated to building and providing infrastructure for the field of digital humanities, high performance computing (HPC) is gaining ground quickly. With the arrival of large language models (LLMs), the motivation for providing HPC infrastructure increased decisively. As a consequence, the first HPC service was established, easing the way for developments in several other areas where access to HPC is needed for enabling solutions otherwise not feasible. Examples of HPC use in Text+ are the Text+ LLM service and the NLP tool MONAPipe.

OpenGPT-X has set a goal to create and train open large language models (LLM) for European languages. Existing language models focus primarily on the English language, and hence perform unfavourably when used for any of the other commonly spoken European languages.
From large-scale benchmarking of multilingual LLMs to introducing Teuken-7B models, our research uncovers how tokenization and balanced datasets enhance cross-lingual performance. Join us in exploring transparent and reproducible innovations shaping the future of multilingual AI.

For the simulation of turbulent dispersed liquid-liquid flows at large scales, coalescence and breakup of droplets is approximated with sub-grid scale closures. For these closures, the root mean square (RMS) droplet fluctuation velocity Urms,d is a decisive input quantity. Recently, Solsvik & Jakobsen [1] proposed an enhanced model to predict Urms,d, which has not been verified yet. Hence, Direct Numerical Simulations (DNS) together with a Volume-of-Fluid (VoF) approach were employed to study the motion of single droplets in a Forced Homogeneous Isotropic Turbulent (FHIT) flow. A parameter study was conducted to investigate the effect of the initial droplet diameter D on Urms,d, and the DNS results were used to assess the model from [1].

Large Language Models (LLMs) are continuously being developed and improved, and there is no shortage of benchmarks that quantify how well they work; LLM benchmarking is indeed a long-standing practice especially in the NLP research community. However, the majority of these benchmarks are not designed for German-language LLMs. We assembled a broad Natural Language Understanding benchmark suite for the German language and evaluated a wide array of existing German-capable models.
This allows us to comprehensively chart the landscape of German LLMs.

Every year we, as the Formula Student Team of RWTH Aachen University, develop a completely new electric race car and revise a previous car to be able to drive autonomously. For our Aerodynamics team, the electric vehicle is the main focus. We try to find the best geometries for our car within the regulatory constraints and while keeping performance compromises with other design areas in mind.

Shallow water simulations are important for climate models, flood or tsunami predictions and other applications. Performing such simulations on unstructured meshes with the Discontinuous Galerkin method is numerically attractive, but a performance challenge on conventional architectures. With a customized dataflow architecture implemented on FPGAs, we have improved performance and power efficiency on a single FPGA and achieved promising initial results when scaling to multiple FPGAs via direct FPGA-to-FPGA interconnects.

In this project the goal is to use algorithms to predict classes of the library classification system “Basisklassifikation” (which can be translated as basic classification). A library classification system is a taxonomy of predefined classes that represent disciplines, subdisciplines, themes or types of publications. Subject librarians assign one or more of these classes to each publication, allowing both final users or retrieval system to use this annotated information for finding publications. As input data we observe mainly bibliographic data, such as for example the title, the name of the publisher, the year of publication and the language of the publication. The algorithms should suggest several classes, which are then analyzed by