PNNL

Abstract

Large-eddy simulation (LES) has been extensively used for atmospheric turbulence, but the traditional approach of using the Monin-Obukhov similarity theory (MO) leads to significant errors in natural convection. In this study, we propose an alternative approach using feedforward neural networks (FNN) trained with direct numerical simulation (DNS) data. To evaluate the performance of our approach, we conduct both {\it a priori} and {\it a posteriori} tests. In the {\it a priori} tests, we compare the model outputs with those from the filtered DNS. We also explore the importance of various input features using the Shapley additive explanations value and conditional average of the filter grid cells. In the {\it a posteriori} tests, we implement the trained FNN models in the System for Atmospheric Modeling (SAM) LES and compare the statistics of surface shear stress and heat flux with those in the DNS. We show that vertical velocity, a traditionally ignored flow quantity, is one of the most important input features in determining the wall fluxes. Increasing the number of input features improves the {\it a priori} test results, but it does not always improve the {\it a posteriori} test results because of the differences between the LES quantities and the filtered DNS quantities. Last, we show that physics-aware FNN models trained with logarithmic and scaled parameters can extrapolate well, whereas those trained with primitive flow quantities cannot.