Tightly Coupling a Machine-Learned Closure Model with a GPU-Based, Computational Fluid Dynamics Solver
DescriptionIn this work, we explore the use of an artificial neural network (ANN) for the prediction of the sub-grid stress (SGS) terms which are needed in the closure of large eddy simulations (LES). We propose an in-situ framework to train and deploy the ANN which predicts components of the SGS tensor using flow-field velocity components. The ANN is trained in on online, in-situ fashion using data from sub-sampled direct numerical simulation (DNS) snapshots of a fully-resolved turbulent flow-field. We instrument NekRS, the high-order CFD code, with the Python interpreter which allows the user to extend their existing code with the capability of using Python frameworks (e.g. CuPY and TensorFlow). We provide details of implementing this approach on GPU-based devices without moving data to the host (i.e. zero copy). The trained framework is assessed for prediction fidelity for a-priori and a-posteriori applications. Comparisons with well-established modeling techniques (e.g. the Smagorinsky model) are made to assess accuracy.
TimeMonday, June 2615:30 - 16:00 CEST
Computer Science, Machine Learning, and Applied Mathematics