A critical review of methods used to assess the uncertainty of air quality models

Most air quality models provide estimates of ensemble mean concentrations, although a few (e.g., SCIPUFF) also provide estimates of the probability distribution function (pdf) for some components of uncertainty, such as the in-plume fluctuations in concentration. There is increasing interest in knowing and communicating the uncertainty of the air quality model predictions, and several recent studies of model uncertainty have appeared in the literature. However, a variety of approaches have been used, with different assumptions and definitions and interpretations. For example, the current author has carried out three Monte Carlo uncertainty studies with regional ozone models, where the effects of uncertainties in model inputs and parameters are accounted for. Others have used the ensemble method, where one model is run for a few variable choices of input conditions or science algorithms, or several different models are run, and the “spread” of the predicted plumes is considered to be a measure of the uncertainty. Still others have carried out a variety of types of sensitivity analysis, where the responses of the predicted concentration to small perturbations in one or more input variables are considered. Because of the increasing use of model systems where a meteorological model (prognostic or diagnostic) is directly linked to an air quality model, the uncertainty analysis has been made more difficult because of the need for the meteorological model to satisfy mass conservation (that is, it is not possible to perturb a single wind speed input to a prognostic model, because the entire model must be rerun to conserve mass). The purpose of the current paper is to critically review the available methods for estimating uncertainty in air quality models, and recommend some promising approaches for the future.