A Genetic Algorithm Method to Assimilate Sensor Data for a Toxic Contaminant Release

Abstract

Following a toxic contaminant release, either accidental or intentional, predicting the transport and dispersion of the contaminant becomes a critical problem for Homeland Defense and DoD agencies. To produce accurate predictions requires characterizing both the source of hazardous material and the local meteorological conditions. Decision makers use information on contaminant source location and transport prediction to decide on the best methods to mitigate and prevent effects. The problem has both observational and computational aspects. Field monitors are likely to be used to detect the release and measure concentrations of the toxic material. Algorithms are then required to invert the problem in order to infer the characteristics of the source and the local meteorology. Here, a genetic algorithm is coupled with transport and dispersion models to assimilate sensor data in order to characterize emission sources and the wind vector. The parameters computed include two dimensional source location, amount of the release, and wind direction. This methodology is demonstrated for a basic Gaussian plume dispersion model and verified in the context of an identical twin numerical experiment, in which synthetic dispersion data is created with the same dispersion model. Error bounds are set using Monte Carlo techniques and robustness assessed by adding white noise. Algorithm speed is tuned by optimizing the parameters of the genetic algorithm. Specific GA configurations and cost function formulations are discussed.