Tomi presented a seminar on “What is parameterization uncertainty and how to evaluate it for ensemble prediction?”
It is widely accepted that parameterizations of sub-grid scale processes in Numerical Weather Prediction (NWP) models are uncertain and that this leads to forecast errors. The impacts of parameterization uncertainty have been demonstrated by numerous sensitivity studies in which either parameter values within a single parameterization were varied or different parameterizations were used with the same NWP model, assuming without justification that the resulting variations of the parameterized processes represent possible and equally valid realizations of the true processes. Although useful for understanding the potential forecast uncertainty such studies do not provide evidence of the parameterization uncertainty itself nor information about its properties. It follows that despite common acknowledgment of the existence of the parameterization uncertainties and the significance of their impacts, vey little is actually known about them. This deficiency could lead to significant errors in ensemble prediction that is meant to combat the model uncertainty, because ad hoc variations of the parameterization in the ensemble may not be representative of the actual uncertainty.
In this seminar results will be presented from a sequence of several studies where an objective nonlinear estimation method was employed to evaluate the parameterization uncertainty using observations. The studies were focused on the microphysics parameterization problem for which both the epistemic and aleatory uncertainty were estimated using satellite microwave and radar reflectivity type of observations. The epistemic uncertainty is represented in terms of variations of physical parameters, whereas the aleatory accounts for stochastic contribution of the processes. The studies were performed with a simplified cloud-resolving model to enable use of fully nonlinear and accurate Markov Chain Monte Carlo (MCMC) estimation method. It is shown that the aleatory uncertainty is more suitable for the ensemble prediction because it leads to unbiased and less uncertain forecast. The potential to estimate the parameterization uncertainty using the Ensemble Kalman Filter (EnKF) data assimilation approach was also tested for both types of uncertainty using the same model and observations.
An audio recording of the presentation is available on the anonymous ftp site:
and a copy of the presentation is also available at: