On the application of an observations-based machine learning parameterization of surface layer fluxes within an atmospheric large-eddy simulation model

Recently, machine learning techniques have been employed to develop improved models for predicting surface-layer fluxes of momentum, heat and moisture based on field observations. Herein we explore refinement to these models, in particular artificial neural networks (NN), and investigate their applicability within an atmospheric large-eddy simulation model as an alternative to the widely adopted standard of Monin - Obukhov (MO) similarity theory. Atmospheric boundary layer (ABL) simulations under different stability conditions are carried out for a variety of scenarios of increasing complexity, from dry steady neutral boundary layers to moist diurnal cycle. Simulations using the NN models result in predicted flux differences with respect to the corresponding MO simulations that are consistent with NN model skill in predicting the tower observations. These differences lead to notable modifications of mean and turbulence quantities throughout the ABL. While these NN models provide an alternative to improve upon MO, it is demonstrated that thorough scrutiny in design and evaluation is necessary to prevent unphysical predictions and establish a robust and generalizable parameterization. Design aspects considered include input feature engineering, applicability under inputs obtained from different heights, biased predictions due to climatological fingerprints, and sensitivity to the choice of activation function. In this context, it is shown how an atmospheric model can contribute toward efficiently investigating these relevant aspects, including expanding the training data set to generalize the NN model to a range of surface roughness values. Finally, we outline remaining challenges to be addressed toward developing a universal parameterization for surface-layer fluxes using machine learning techniques.