Building-block-flow model for large-eddy simulation: GPU implementation and application to the CRM-HL

The Aerospace Computational Design Lab and Computation Turbulence Group present a GPU-based version of the Building-block Flow model (BFM) for wall-modeled large-eddy simulation (WMLES). The researchers evaluated how the model’s cost scales with the number of neurons and layers, with results showing that for the coarse grid resolutions considered, BFM offers small improvements or similar performance with respect to other models that were studied.

Authors: Yuenong Ling, Samuel Costa, Gonzalo Arranz, Konrad Goc, Mory Mani and Adrian Lozano-Duran
Citation: AIAA 2024-3686. AIAA Aviation Forum, ASCEND July 2024

Abstract:
We present the GPU-based implementation of the Building-block Flow model (BFM) for wall-modeled large-eddy simulation (WMLES). The BFM is an artificial neural network (ANN) closure model that unifies subgrid-scale (SGS) and wall models for WMLES. It locally represents the flow as a collection of simple flows that encapsulate the essential physics to model the under resolved scales in WMLES. Previous implementations of BFM were for CPU-based solvers. Here, we discuss the porting of the BFM to GPU architectures. The implementation is conducted in CUDA. We also investigate the scaling of the model cost as a function of the number of neurons and layers. The computational cost of GPU-based BFM is benchmarked against WMLES with the GPU-based versions of Vreman and Dynamic Smagorinsky models (DSM), each coupled with the equilibrium wall model (EQWM). The computational cost of the full WMLES solver with BFM is 7.3% slower than Vreman and 5.0% faster than DSM. We also test the BFM in the Common Research Model (CRM-HL) for the 5th High Lift Prediction Workshop (HLPW5). Our results show that for the coarse grid resolutions considered (∼ 35 million control volumes), BFM offers small improvements or similar performance with respect to DSM with EQWM.