- Kopplung turbulenter Grenzschichtsimulation mit einem Transformer
Hilgers, Tom; Müller, Matthias S. (Thesis advisor); Meinke, Matthias (Thesis advisor); Orland, Fabian (Consultant); Hübenthal, Fabian (Consultant)
Aachen : RWTH Aachen University (2026)
Master Thesis
Masterarbeit, RWTH Aachen University, 2025
Abstract
Ongoing efforts to reduce greenhouse gas emissions in the aviation sector include research on Active Drag Reduction (ADR) methods. However, evaluating ADR methods requires computationally expensive Computational Fluid Dynamics (CFD) simulations of the drag-inducing Turbulent Boundary Layer (TBL). This high cost has motivated recent efforts to accelerate the simulations using Machine Learning (ML). Prior works introduced a Transformer model for this task, proposed a prototype coupling to the multi-physics solver m-AIA and developed a modular Fortran-based connecting interface. In this thesis, we implement a complete coupling between m-AIA and the Transformer by redeveloping the Fortran-based interface in C++, thereby enabling a flexible connection of modern CFD solvers with ML models interchangeably through either the PhyDLL or AIxeleratorService coupling libraries. With both libraries we achieve non-stagnating scalability and high parallel efficiency, while the AIxeleratorService yields the lowest absolute runtimes. To maximize efficiency on modern heterogenous clusters, we extend the AIxeleratorService to distribute inference workloads across available CPU and GPU resources. We derive a formal model to determine the optimal work distribution and validate our approach experimentally, yielding near optimal results, with performance and energy efficiency gains that scale with the number of allocated CPUs to GPUs. We also conduct a holistic assessment of the Transformer for its intended task. We find that low conventional errors mask substantial physical inaccuracies, which negatively affect the coupled solver solution. Our architectural analysis identifies fundamental limitations and proposes improvements as well as alternative models. To accelerate future development, we implement a modular training pipeline and successfully demonstrate its capabilities by training a prototype model.
Institutions
- Chair of High Performance Computing (Computer Science 12) [123010]