This talk will present geoinformatics algorithms based on machine learning, evolutionary computation and high performance computing. The first problem domain relates to the source characterization of an unknown atmospheric release. The typical scenario consists in a set of ground or remote sensing measurements detecting concentrations of a contaminant. The goal is to reconstruct the characteristics of the source using a transport and dispersion model driven by machine learning. At each iteration, dispersion simulations are run from each candidate source, and the simulated concentrations are compared to the observed measurements. The goal is to minimize the error between the simulations and the observations. Results are shown using data from the Prairie Grass experiment, Dipole26 experiment and for the 2011 Fukushima nuclear accident.