Efficient LES Parametric Studies via ANN-Based Multi-Fidelity Modeling and Adaptive Sampling

Abstract

Large Eddy Simulations (LES) are increasingly used in industry due to their superior accuracy compared to traditional statistical methods like Reynolds-Averaged Navier-Stokes (RANS) simulation. However, their high computational cost remains a major obstacle to performing daily parametric studies in engineering design offices. The objective of this work is to improve the efficiency of LES-based parametric studies through multi-fidelity surrogate modeling. Taking into account the computational cost of each turbulence modeling approach, multi-fidelity technic propose to combine limited number of LES results with more numerous RANS simulations. To achieve this, we use Artificial Neural Networks (ANN), which are particularly effective at capturing complex relationships between fidelity levels and handling discontinuities. To further reduce computational cost, we propose a new adaptive sampling strategy that selects high-fidelity LES points based on an estimation of interpolation error. This approach enhances the accuracy of the multi-fidelity method by efficiently allocating computational resources where they are most needed. The proposed strategy is first validated on an analytical test case before being applied to the study of the lift coefficient as a function of the angle of attack for a NACA0012 airfoil. We demonstrate that with only five LES evaluations, our method accurately captures the main features of this function, including the stall angle. This work paves the way for more efficient LES-based parametric studies.