Both inadvertent and intentional releases of airborne toxins pose a threat to society. Managing this threat requires realtime prediction of time evolving contaminant concentration fields from limited observations. These data assimilation algorithms perform best when provided with numerous observations within the drifting contaminant plume. Yet, the sensor density required to achieve this many “hits” early in the plume's lifetime can be economically impractical. Thus, solution of the contaminant puff prediction problem requires mobile sensors capable of intercepting and mapping a puff once it has been detected by a sparse array of fixed sensors.

Realtime contaminant puff interception becomes more feasible and less labor intensive if the unmanned aeronautical vehicles (UAV) involved are also autonomous, i.e. able to make their own routing decisions. This interception problem is nonlinear and can be solved either analytically or by iterative optimization. In contrast, the subsequent puff mapping problem has no analytic solution. Thus, while a neural network approach can be applied to both problems, the first can be handled via supervised learning while the second requires unsupervised learning. Experiments with the interception problem demonstrate that neural network performance nears 100% only when the problem is geometrically transformed to remove a discontinuity in the function relating UAV course to intercept success. Training a neural net to handle the post-intercept puff mapping problem requires that the traditional back propagation approach be replaced by a method that does not require a priori knowledge of the optimal solution. One way to achieve this unsupervised training is via a genetic algorithm.