Invariance-embedded Machine Learning Sub-grid-scale Stress Models for Meso-scale Hurricane Boundary Layer Flow Simulation I: Model Development and $\textit{a priori}$ Studies

Abstract

This study develops invariance-embedded machine learning sub-grid-scale (SGS) stress models admitting turbulence kinetic energy (TKE) backscatter towards more accurate large eddy simulation (LES) of meso-scale turbulent hurricane boundary layer flows. The new machine learning SGS model consists of two parts: a classification model used to distinguish regions with either strong energy cascade or energy backscatter from those with mild TKE transfer and a regression model used to calculate SGS stresses in regions with strong TKE transfer. To ease model implementation in computational fluid dynamics (CFD) solvers, the Smagorinsky model with a signed coefficient $C_s$, where a positive value indicates energy cascade while a negative one indicates energy backscatter, is employed as the carrier of the machine learning model. To improve its robustness and generality, both physical invariance and geometric invariance features of turbulent flows are embedded into the model input for classification and regression, and the signed Smagorinsky model coefficient is used as the output of the regression model. Different machine-learning methods and input setups have been used to test the classification model's performance. The F1-scores, which measure balanced precision and recall of a model, of the classification models with physical and geometric invariance embedded can be improved by about $17\%$ over those without considering geometric invariance. Regression models based on ensemble neural networks have demonstrated superior performance in predicting the signed Smagorinsky model coefficient, exceeding that of the dynamic Smagorinsky model in $\textit{a priori}$ tests.