Orland, Fabian; Müller, Matthias S. (Thesis advisor); Bischof, Christian (Thesis advisor)
Aachen : RWTH Aachen University (2026)
Dissertation / PhD Thesis
Dissertation, RWTH Aachen University, 2026
Abstract
The field of computational fluid dynamics (CFD) is concerned with the numerical simulation of turbulent or reactive flow problems that occur in important real world applications. These simulations require powerful high-performance computing (HPC) systems to be carried out efficiently in reasonable time. However, large scale simulation cases of practical industrial relevance may still exceed the capabilities of even today's most powerful supercomputers. In these cases, special modeling techniques to accurately describe the relation between different physical quantities of interest instead of simulating the real physics are employed to reduce the computational effort down to a feasible level. Due to the availability of huge datasets and powerful hardware, data-driven models from deep learning (DL) gained a lot of traction in recent years. The investigation of deep learning-based physical modeling approaches requires the coupling of existing CFD solvers with the inference of new DL models like deep neural networks, for example. While many traditional CFD solvers have been optimized over decades to run efficiently on CPU-based architectures, the inference of DL models can be significantly accelerated by GPUs. From a computational perspective, the desired coupling is challenging because modern heterogeneous HPC systems comprising multicore CPUs and massively parallel GPUs need to be exploited efficiently. This thesis investigates the coupling of highly parallel CFD simulation with deep learning and provides a novel scalable general coupling method to efficiently accelerate the inference task on multiple distributed GPUs while keeping the CFD solver execution on CPUs. Four real world coupled CFD+DL use cases are analyzed to derive application requirements based on which a formal abstract coupling model is defined to understand the semantics of a general coupling method, that supports a wide range of existing CFD solvers and DL model architectures. Moreover, a new hybrid inference method is proposed to optimize the utilization of heterogeneous hardware resources. The flexibility and expandability of the general coupling method is demonstrated by an extension to also support online training of DL models coupled to a CFD solver. The developed general coupling method and both extensions are implemented in a new deep learning coupling library including the new AIxeleratorService library as a distributed inference backend. The computational performance of the implemented coupling method is evaluated using all four use cases considered in this thesis to demonstrate the applicability to real world applications with efficient scalability.
Institutions
- Chair of High Performance Computing (Computer Science 12) [123010]