Despite our increased understanding in the relevant physical processes, forecasting radiative cold pools and their associated meteorological phenomena (e.g. fog and freezing rain) remains a challenging problem in mesoscale models. In this thesis we recognize that our current modeling approach is flawed and incomplete, rendering us unable to forecast these high impact events. Using the Weather Research and Forecasting (WRF) model, fundamental deficiencies are exploited and new physical parameterizations are introduced to address these issues without degrading the model in other areas and time.

It was found that the default model diffusion, which is calculated on model sigma coordinates, in addition to the 6th order numerical filter, prevented the formation of cold pools. Furthermore, soil moisture in and around valleys from both natural and anthropogenic sources, the vertical resolution, and model physics were all found to be important and play at least some role in forecasting these events. Finally, key physical processes governing the evolution and life cycle of cold pools were missing and subsequently introduced to a boundary layer parameterization to substantially improve its forecasting ability. Through the knowledge presented here, this and other key modifications have accomplished a numerical weather model of more general use and applicability which will not only be of use to the modeling community but society at large.

Using the Weather Research and Forecasting (WRF) model, an intra-model physics ensemble was created for July 2005 in hopes of finding a physics configuration that could most accurately recreate the typical summertime conditions of Southern California. The ensemble consisted of twenty-four members that mixed different Land Surface Models (LSM) with various Planetary Boundary Layer (PBL) parameterizations currently offered in WRF. It was found that the Pleim-Xiu LSM coupled with either the asymmetric convective model version 2 or the Yonsei University PBL scheme created the most skillful ensemble member.

In addition to the physics ensemble, short integrations along with different initializations proved to be beneficial. Initializing the model off of the North American Regional Reanalysis (NARR) instead of the North American Mesoscale (NAM) model was, in this case, extremely useful. The most accurate simulation of July 2005 was found when taking integration methods, initializations, and model physics into consideration.