EN: Fluid flow phenomena play an important role across a spectrum of disciplines in science and engineering. Since experiments are not always an efficient or feasible way to study them, computer simulations of these phenomena have turned out to be one of the essential tools for progress in this area. However, computational fluid dynamics is a challenging field, especially when characteristics on multiple length and time scales need to be captured. Coupled methods are one way to simulate such a flow. They constitute a compromise between the physical accuracy of a fine-scale method and the computational efficiency of a coarse-scale method, providing a way to combine the best aspects of both methods. Coupling, in a domain decomposition sense, allows fine-scale details to be present only in the small region of interest where they are needed, while the coarse-scale method allows the modeling of a larger domain. Basic coupled methods and software come with many limitations, since inconsistent or noisy coupling data and the required computational resources restrict the applicability. The simulation execution time is the most critical aspect that needs to be enhanced in order to accelerate the availability of results and generate new scientific insights. Thus, the utilization of high-performance computing systems, adequate parallelization methods and well-chosen software designs are often inevitable. However, when aiming to benefit from recent and upcoming computing systems with a large number of cores, the scalability of the simulation methods limits the reachable execution times. Scalable performance enhancements hence offer the scientific value of enabling emerging future applications which would otherwise be infeasible. One example of coupled methods for multiscale flows are molecular-continuum simulations, where the fine-scale method is a molecular dynamics system, i.e. a particle-based solver that explicitly determines individual trajectories based on molecule interaction potentials. A molecular-continuum simulation software is given by MaMiCo, an open source coupling framework with modular software design for massively parallel execution on high-performance computing systems. Prior to the developments in the scope of this thesis, MaMiCo does not incorporate coupling data-driven enhancements such as advanced noise filters, machine learning surrogates, or temporal parallelization. This thesis presents extensions of MaMiCo by new features, leading to three novel ways to enhance multiscale flow simulations based on data-driven analysis of the coupling data. First, thermal noise filtering methods make the coupling data more consistent across models on different scales. A novel space-time formulation of the non-local means algorithm is introduced, outperforming conventional filtering methods. Second, machine learning-based surrogates reduce computational cost of expensive fine-scale solvers. An advanced hybrid model architecture is presented where a convolutional autoencoder deals with the spatial extent of the flow data, while a recurrent neural network is used to capture its temporal correlation. Third, parallel-in-time integration allows to expand a coupled simulation to multiple temporal scales. An implementation based on a variant of the Parareal algorithm is presented here, where a Lattice Boltzmann solver is used as hydrodynamic predictor to supervise the microscopic system in time. The capabilities of the novel data-driven enhancements are demonstrated in molecular-continuum flow scenarios including a three-dimensional Couette flow and a vortex street scenario. Large-scale MaMiCo scalability experiments on supercomputing platforms validate the benefits of the presented new methods.
Piet Jarmatz (2026) studied this question.