Shear-Convection flux decomposition in atmospheric boundary layer using neural network

<div>&#160;</div><div>The representation of turbulent mixing in climate models is challenging due to eddies' wide range of scales and modeling limitations thus <div>turbulent flux parameterization remains one major source of climate models inaccuracy. Most turbulent flux parameterizations decompose total flux to a diffusion term, which models the local small scale mixing and a non diffusive term&#160; (called non-local) that represents the contribution from coherent structure. However, these two terms are not coupled and it is not clear how much each term should contribute to the total flux.</div> </div><div>Given the availability of high resolution numerical simulations, one promising approach for more accurate representation of turbulent fluxes in climate models is the use of machine-learning (ML) based parameterization. The high resolution simulations such large eddy simulations (LES) have substantially less biases so that it provides a rich dataset for training ML models that have the potential to outperform conventional parameterizations. The use of neural networks in atmospheric science has proven promising due to its ability in emulating nonlinear physical systems.</div><div>&#160;In this work, we use a novel neural network architecture that combines variational auto-encoders (VAE), a nonlinear dimensionality&#160; reduction technique, with an encoder-decoder network to parameterize the turbulent fluxes. VAE maps high dimensional information into a very compact representation which is then used to predict turbulent fluxes. Our neural network is trained on coarse grained high resolution LES simulation of dry convective boundary layer. We show that by enforcing few physical constraints in the latent space of VAE, we can decompose scalar turbulent fluxes into two main modes of variability. These two modes represent the contribution of convection and shear to scalar turbulent flux. This finding can be used to inform conventional parameterization of turbulent fluxes.</div>