(RP07) Towards Clean Propulsion with Synthetic Fuels: A Cluster-Modularized Approach Employing Hierarchies of Simulations and Deep Learning
AI/Machine Learning/Deep Learning
Clouds and Distributed Computing
TimeTuesday, June 18th8:30am - 10am
DescriptionDevelopment of new clean propulsion technology is difficult as the parameter space of possible realizations is large and cannot be investigated cost efficiently by experimental methods as, for example, geometry optimizations require manufacturing. The only way to overcome this issue is by combining experimental and simulation techniques. However, also simulation techniques struggle with the wide range of involved scales, and a direct approach would exceed the currently available computing capacities. Therefore, this work presents a hierarchical simulation approach employing high-order models (DNS) as well as reduced-order models (LES) utilized on different German supercomputer. While the used high-order models naturally rely on all-to-all communication, the reduced-order models mostly require local communication. Consequently, the DNS were performed on JUQUEEN with its 5D torus network and the LES on Hazel Hen. As the data of the DNS are required as BC for the LES, coupling is needed. In this work, a CNN was trained with the DNS data in order to reduce the required data transfer from TBs to GBs. This training was done using modularized computing on JUQUEEN/JURECA with on-the-fly data streaming. In this way, JURECA’s GPUs were efficiently used and classical I/O avoided. Pre- and postprocessing was done on CLAIX16. The optimal cluster choices enabled very good scaling and node performance. Overall, the combination of hierarchical simulations, modularized computing and deep learning allowed an iterative technology development cycle, as feedback to the experiment was possible on time (edge2cloud). Optimizations with respect to nozzle geometry, injection conditions and fuel properties were realized.