Presentation
PhyDLL - Physics Deep Learning CoupLer: An Open-Source High-Performance Coupling Library
Presenter
DescriptionThere has been recently a significant shift towards using data-driven approaches, particularly deep learning (DL) in computational fluid dynamics (CFD). While these techniques have shown great promise in improving the accuracy of fluid simulations, there is a growing need to consider the high-performance efficiency and robustness when solvers are querying online DL inference. To address these concerns, the open-source coupling library PhyDLL (Physical Deep Learning coupLer) has been developed. It allows high-performance data transfer and processing between massively-parallel fluid solvers and distributed DL inferences. This library includes different coupling schemes that fit the context encompassing the CFD solver and DL engine. PhyDLL provides therefore a Fortran interface, that cater to a large portion of CFD solvers, as well as a Python one which aligns well with prime DL libraries such as Tensorflow and Pytorch. Toward exascale, PhyDLL is well designed to take advantage of computing capabilities of hybrid architecture (CPU-GPU) of modern clusters. A C/C++ interface will also be available for better portability and to reach an even wider range of users. Finally, significant physical and performance results have been achieved using PhyDLL to couple the CFD solver AVBP to graph and convolutional neural network for combustion and aerodynamic use-cases.
TimeWednesday, June 2814:30 - 15:00 CEST
LocationWisshorn
Session Chair
Event Type
Minisymposium
Physics
