Implementation and Validation of an Offshore-specific Machine Learning Surface Layer Model in WRF

Prediction of surface fluxes – in numerical weather prediction models or for observations without having direct flux measurements – has often been performed using Monin-Obukhov Similarity Theory (MOST). This theory is based on empirically defined stability functions from a few widely accepted field experiments conducted in environments over land and characterized by horizontal homogeneity and stationarity. These assumptions do not hold for coastal and offshore environments.

In this study, offshore observational datasets (including FINO-1, MVCO-ASIT, and CASPER-WEST) that contain near surface flux measurements are used to train machine learning surface layer (MLSL) models to estimate sensible heat flux and momentum flux. The ML models, random forest and artificial neural networks, are not constrained by the simple regression techniques that are used in MOST and are instead able to explore new, non-linear relationships between various features including ocean surface variables. These models are then implemented in the Weather Research and Forecasting (WRF) model. Performance is then compared between simulations with MOST and the new MLSL models in offshore environments. Due to the dearth of long-term flux measurements over the ocean, generalization of the MLSL models is explored. First, we apply the models in locations other than the observation site. Second, we utilize combined datasets of derived predictors that can be interpreted as site and sensor location independent from multiple observation sites.