A large ensemble approach to quantifying internal model variability within the Weather Research & Forecasting (WRF) numerical model

The Weather Research and Forecasting (WRF) community model is widely used to explore cross‐scale atmospheric features. Although WRF uncertainty studies exist, these usually involve ensembles where different physics options are selected (e.g., the boundary layer scheme) or adjusting individual parameters. Uncertainty from perturbing initial conditions, which generates internal model variability (IMV), has rarely been considered. Moreover, many off‐line WRF research studies generate conclusions based on a single model run without addressing any form of uncertainty. To demonstrate the importance of IMV, or noise, we present a 4‐month case study of summer 2018 over London, UK, using a 244‐member initial condition ensemble. Simply by changing the model start time, a median 2‐m temperature range or IMV of 1.2 °C was found (occasionally exceeding 8 °C). During our analysis, episodes of high and low IMV were found for all variables explored, explained by a relationship with the boundary condition data. Periods of slower wind speed input contained increased IMV, and vice versa, which we hypothesis is related to how strongly the boundary conditions influence the nested region. We also show the importance of IMV effects for the uncertainty of derived variables like the urban heat island, whose median variation in magnitude is 1 °C. Finally, a realistic ensemble size to capture the majority of WRF IMV is also estimated, essential considering the high computational overheads (244 members equaled 140,000 CPU hours). We envisage that highlighting considerable IMV in this repeatable manner will help advance best practices for the WRF and wider regional climate modeling community.