2.7      A machine learning surface layer parameterization for WRF.

Gagne, David, Tyler McCandless, Branko Kosovic, Amy DeCastro, Thomas Brummet, Sue Haupt, Richard Loft, National Center for Atmospheric Research, and Bai Yang, NOAA

Surface layer parameterization schemes estimate the momentum, sensible heat, and latent heat fluxes between the land surface and atmosphere. These schemes are based on Monin-Obukhov similarity theory, which utilizes empirical stability functions to estimate wind, temperature, and moisture near the surface. Because of the limitations of Monin-Obukhov, we developed and evaluated machine learning model estimates of the surface layer fluxes based on long time series of observations. We train random forests and neural networks to predict the scaling parameters needed for flux calculation using data from the Netherlands and Idaho. We show that the machine learning models can produce improved representations of the fluxes compared with MOST at these locations, particularly for sensible and latent heat fluxes. We use machine learning interpretation methods, such as variable importance and partial dependence plots, to understand the input sensitivities for the machine learning models and Monin-Obukhov. We discuss how we developed a Fortran random forest inference module to incorporate the machine learning surface layer models back into WRF. Finally, we will show initial evaluations of WRF single-column runs using the machine learning surface layer scheme.