A New Ensemble of Perturbed-Input-Parameter Simulations by the Community Atmosphere Model

摘要: Uncertainty quantification (UQ) is a fundamental challenge in the numerical simulation of Earth's weather and climate, other complex systems. It entails much more than attaching defensible error bars to predictions: particular it includes assessing low-probability but high-consequence events. To achieve these goals with models containing large number uncertain input parameters, structural uncertainties, etc., raw computational power needed. An automated, self-adapting search possible model configurations also useful. Our UQ initiative at Lawrence Livermore National Laboratory has produced most extensive set date simulations from US Community Atmosphere Model. We are examining output about 3,000 twelve-year climate generated specialized software framework, model's accuracy as function 21 28 parameter values. Most parameters we vary related boundary layer, clouds, sub-grid scale processes. prescribe surface conditions (sea temperatures sea ice amounts) match recent observations. Fully searching this 21+ dimensional space impossible, sensitivity ranking algorithms can identify having relatively little effect on variety fields, either individually or nonlinear combination. Bayesian statistical constraints, employing observations metrics, seem promising. Observational constraints will be important next step our project, which compute interactively, study change due increasing atmospheric carbon dioxide.