Recovering an Event Realization with a Genetic Algorithm

Making a prediction under uncertainty conditions is often required in weather prediction. How can artificial intelligence help to identify and predict a situation given observations of the current state of a chaotic flow? Predicting a particular realization of an evolving flow field requires knowledge of the current state of that field and assimilation of observations into the model. Here we consider the example of modeling atmospheric transport and dispersion of a contaminant when the observation is of the transported contaminant, a problem that exemplifies the issue of turbulent flow. In this case, the problem is compounded by the fact that the field observed is a tracer that is advected and mixed by the flow field, but does not directly alter the flow field. This one-way coupled system presents a challenge: one must first infer the changes in the flow field from observations of the contaminant, then assimilate that to recover both the advecting flow and information on the subgrid processes that provide the mixing. To accomplish such assimilation requires a robust method to match the observed contaminant field to that modeled.

The approach used here is a genetic algorithm used to optimize the variational problem (GA-Var) of matching the modeled flow with that observed. Given contaminant sensor measurements and a transport and dispersion model, one can back-calculate unknown meteorological parameters. In this case, we demonstrate the dynamic recovery of unknown meteorological variables, including the transport variables that comprise the “outer variability” (wind speed and wind direction) and the dispersion variables that form the “inner variability” (contaminant spread). The optimization problem is set up in an Eulerian grid space, where the comparison of the concentration field variable between the predictions and the observations forms the cost function. The dispersion parameters apply to Lagrangian puff space, where one estimates the transport and dispersion parameters. This method is demonstrated in the context of recovering transport and dispersion of a large eddy simulation by using a Gaussian puff dispersion model to match the flow.