Comparison of Several Neural Network-Enhanced Sub-Grid Scale Stress Models for Meso-Scale Hurricane Boundary Layer Flow Simulation

Abstract

The complicated energy cascade and backscatter dynamics present a challenge when studying turbulent flows in storms at the meso-scale. When performing standard large-eddy simulations (LES), sub-grid scale (SGS) stress models usually fail to consider energy backscatter. These models assume that kinetic energy only moves continuously from larger to smaller scales. However, coherent energy backscatter structures exist when analyzing hurricane boundary layer flows at the meso-scale. Our recent research has shown that machine-learning SGS models trained with high-resolution data can effectively forecast forward and backward energy transfers in meso-scale hurricane-like vortex flows. Therein, physical and geometrical invariances were introduced to better represent flow physics. This further improved the predictability and generalizability of machine-learning-enhanced SGS models. In this study, we compare the performance of several machine-learning-enhanced SGS models, especially those based on neural networks (NNs), with varying physical and geometrical invariance embedding levels for SGS stress modeling in an a priori sense, which sets the cornerstone for ongoing a posteriori tests of NN models.