Given an unknown but detected release of a toxic agent, the current NARAC capability for reconstructing source characteristics is a highly manual procedure that often relies on analyst judgement and requires many hours of computations for a refined analysis. There is no automated, optimization approach to estimating the source characteristics. A fast running, prototype atmospheric inversion model has been developed for use as a test bed for the evaluation of source inversion schemes. The model was applied to a simple puff release scenario to test the relationship between the amount of sampled data obtained and the accuracy of the determination of the inverted source parameters. The initial inversion scheme chosen for the test bed model utilizes the Marquardt method coupled to a Gaussian puff atmospheric dispersion model driven by a COAMPS model wind field. The inversion scheme results are used in conjunction with a sensor realization probability model for a sensor realization scenario consisting of lo00 possible sensor realizations. The results of the initial test calculations indicate that the inversion procedure produces good results for the four source parameters, location (x, y), release time, and strength along with reasonably well defined maximum probabilities for the sensor realization scenarios. The model runs relatively fast, taking -100 seconds per inversion on a Sparc 10 workstation.

A technique has been developed to obtain selected source parameters for a release of toxic material into the atmosphere by inversion of a set of sensor observations. The technique utilizes the Marquardt inversion method coupled to a Gaussian puff atmospheric dispersion model. The major objective of this report is to perform a set of sensitivity calculations to explore the robustness of the source location inversion procedure to variations in source location relative to sensor location, spacing of sensors, sensor integration period, and sensor observation times. The results of the tests for variation of source location show that the inversion accuracy is sensitive to source placement relative to a rectangular sensor array with most accurate inversions occurring when the source is aligned with a downwind row of sensors and least accurate when the source is placed between downwind sensor rows. When the density of the sensor array was increased by a factor of 4, i.e. spacing cut in half, the case with the poorest source location inversion accuracy in the first test significantly improved. The results of tests for variation of sensor integration time show that the most accurate source location inversion occurs for the smallest integration time and the least accurate for the largest integration time. The results of tests for variation of the number of observations show that the most accurate source location inversion occurs for the case with the most observation times and the accuracy degrades as the number of observation times decreases. As a general rule throughout all of the sensitivity tests performed here, the most accurate source inversions occur when the number of data points (in both time and space) is maximized.