Evaluation of Surface Fluxes in WRF Using Eddy-Covariance Flux Measurements in the Western and Eastern U.S.

The accurate simulation of surface fluxes in the Weather Research and Forecasting (WRF) Model is highly dependent on surface representation. Land use patterns in the San Joaquin Valley (SJV) of California (CA) and the northeastern United States (US) feature a combination of agricultural landscapes, urbanized areas, and forests. This intricate composition presents a challenge to the model simulation of surface fluxes, subsequently affecting the simulation of the atmospheric boundary layer (ABL), which is central to the development of high-quality air quality regulations. In this study, the WRF model simulations were configured with the physical modules commonly adopted by state regulatory and environmental agencies. Compared to the year-long eddy-covariance flux measurements collected from sixteen sites with various land use types, the Pleim-Xiu land surface model (PX LSM) used in the WRF numerical weather reanalyses of the two regions shows substantial biases. In the SJV of CA, the model overestimates sensible heat (H) fluxes by a factor of two to three and underestimates latent heat (LE) fluxes by 50~100% at irrigated cropland and orchards during spring and summer days. During the same time, at the non-irrigated savannas, the model overestimates H fluxes by 30~50%, and the model bias in LE fluxes is negligible. In the US Northeast, all sites are classified as non-irrigated areas. The model overestimates both H and LE fluxes by less than 100% throughout the year during the days. The most pronounced model biases in H and LE fluxes are observed in the summer at two urban sites and a turfgrass site, respectively. Model biases of surface net radiation (Rnet) are small compared to the heat flux biases at irrigated sites. During the day in all seasons, the model generally underestimates Rnet in the SJV, except for the two non-irrigated savanna sites, while in the US Northeast, the model overestimates Rnet to a similar degree. The biases observed in the Rnet are likely attributed to the model biases in the incoming solar radiation (SWin). The diurnal cycles of SWin residuals show similar patterns and magnitudes to the Rnet residuals over irrigated orchards in the SJV and non-irrigated croplands and forests in the US Northeast. Different shortwave radiation schemes were applied in the two regions, and the model biases in SWin shifted from underestimation in the SJV to overestimation in the US Northeast, aligning with the model biases in Rnet. The simulations of momentum flux do not show discernible diurnal, seasonal, or annual patterns across regions or sites. Unlike heat or radiative fluxes, the model bias in momentum flux is comparable in magnitude between day and night and among seasons. The absence of accurate representation of irrigation in the WRF simulations appears to cause large biases in the summer daytime turbulent heat fluxes in the SJV. Northeastern US region flux biases are much smaller in magnitude. Future work will test the impact of these flux biases on boundary layer development and the potential for assimilation of satellite remote sensing to minimize these biases.