Automatic generation of large ensembles for air quality forecasting using the Polyphemus system

Abstract

This paper describes a method to automatically generate a large ensemble of air quality simulations. Such an ensemble may be useful for quantifying uncertainty, improving forecasts, evaluating risks, identifying process weaknesses, etc. The objective is to take into account all sources of uncertainty: input data, physical formulation and numerical formulation. The leading idea is to build different chemistry-transport models in the same framework, so that the ensemble generation can be fully controlled. Large ensembles can be generated with a Monte Carlo simulations that address at the same time the uncertainties in the input data and in the model formulation. This is achieved using the Polyphemus system, which is flexible enough to build various different models. The system offers a wide range of options in the construction of a model: many physical parameterizations, several numerical schemes and different input data can be combined. In addition, input data can be perturbed. In this paper, some 30 alternatives are available for the generation of a model. For each alternative, the options are given a probability, based on how reliable they are supposed to be. Each model of the ensemble is defined by randomly selecting one option per alternative. In order to decrease the computational load, as many computations as possible are shared by the models of the ensemble. As an example, an ensemble of 101 photochemical models is generated and run for the year 2001 over Europe. The models' performance is quickly reviewed, and the ensemble structure is analyzed. We found a strong diversity in the results of the models and a wide spread of the ensemble. It is noteworthy that many models turn out to be the best model in some regions and some dates.