Machine Learning Parameterization of the Surface Layer: Integration with WRF

The surface layer is the interface between the atmosphere and land surface where the exchange of momentum, sensible heat, and latent heat occurs. All atmospheric models parameterize the effects of the surface layer using Monin-Obukhov similarity theory, which assumes logarithmic wind, temperature, and moisture profiles for the surface layer that are adjusted through the use of empirically-derived stability functions. In order to address the limitations of similarity theory, we have developed a machine learning surface layer parameterization and implemented it within the WRF Single Column Model. We trained random forests on observations from sites in the Netherlands and Idaho. The random forest models predict the friction velocity, temperature scale, and moisture scale for a given atmospheric profile. These parameters are used to derive the momentum, sensible heat, and latent heat fluxes as well as the heat and moisture exchange coefficients used to force the land surface model. In offline comparisons with observations at both Cabauw and Idaho, the random forest model outperforms similarity theory in estimating fluxes, even when training the model in one location and applying to the other. In WRF SCM runs, the random forest surface layer parameterization generally reproduces the diurnal cycle of the boundary layer for both temperature and moisture. We will discuss how different data processing strategies affect the machine learning model sensitivities and the physical consistency of the WRF single column model runs. We use machine learning interpretation techniques, such as partial dependence plots and permutation variable importance, to investigate the sensitivities of both the random forests and similarity theory. We will also discuss how the machine learning parameterizations perform when paired with different land surface model and boundary layer scheme options.