Despite many recent improvements, climate models continue to poorly simulate extreme precipitation. I attempted to improve prediction of extreme precipitation, focusing on daily 95th percentile (P95) events, and to better understand the source of model biases in three ways: 1) determine which physics processes P95 is most sensitive to and which parameterization schemes best represent these processes; 2) understand the underlying mechanisms through which these processes impact P95; and 3) maximize advantages from the ensemble of the best performing models.

First, to determine the sensitive processes affecting P95, I tested a 25-member ensemble of different physics configurations in the regional Climate-Weather Research and Forecasting model (CWRF) for 36-yr historical U.S. simulations. Of these, P95 simulation was most sensitive to cumulus parameterization. Overall, the ensemble cumulus parameterization best represented P95 seasonal mean spatial patterns and interannual variations, while one traditional cumulus scheme generally overestimated P95 and the other three severely underestimated P95, especially over the Gulf States (GS) and the Central-Midwest States (CM) in convection-dominated seasons.

Second, I built structural equation models (SEMs) to identify the underlying processes through which cumulus parameterization affects precipitation. I discovered five distinct physical mechanisms, each involving unique interplays among water and energy supplies and surface and cloud forcings. The relative importance of these factors varied significantly by season and region. For example, water supply is the dominant factor for P95 in CM, but its effect reversed from positive in summer to negative in winter due to changes in the prevailing precipitation system. In contrast, the predominant factors affecting P95 in GS were cloud forcing in summer, but surface forcing in winter. Since the choice of cumulus parameterization affected how water and energy supplies acted through surface and cloud forcings, it determined CWRF’s ability to simulate extreme precipitation.

Third, I improved P95 prediction by developing an optimized multi-model ensemble based on the Bayesian Model Averaging (BMA) approach. BMA is a model-selection method that weights ensemble members to create an optimal composite. However, many BMA methods rely on maximum likelihood estimation and thus may be flawed when the true solution is not among the ensemble, as is the case in extreme precipitation. To resolve this issue, I adapted three BMA variations to fit the needs of extreme precipitation problems. These methods significantly improved performance compared to both the ensemble mean and the single best model and provided a more reliable confidence interval.

My work shows that to improve extreme precipitation simulation, a better understanding of physics processes, especially cumulus processes, is critical. For this, I applied the SEM framework, for the first time in the climate community, to uncover the underlying physical mechanisms essential to regional extreme precipitation predictions. Furthermore, I adapted new BMA methods into extreme precipitation ensembles to maximize the benefits from the most physically advanced models. These advances may help improve the prediction of extreme precipitation occurrences and future changes, one of the most difficult modeling challenges and one with huge socioeconomic significance.