Machine Learning Reconstruction of Atmospheric Releases from Catastrophic Radiological Events

Nuclear power accidents, nuclear explosions, and other catastrophic events can release hazardous levels of radiological materials to the atmosphere that disperse and pose a threat to humans and the biota. Under these circumstances it may be difficult or dangerous to acquire direct measurements to constrain the release characteristics of an event. Machine learning techniques that combine sensor data with atmospheric simulations offer an alternate strategy for reconstructing different aspects of radiological releases, including emission rates, source geometries and locations, and other factors. Once trained, computationally efficient machine learning algorithms can also be used to evaluate millions of radiological release scenarios to estimate probability density functions of a particular event. Using machine learning approaches, we have reconstructed the release characteristics of the Fukushima nuclear disaster, nuclear reprocessing accidents, and other radiological events. We present an overview of the machine learning source reconstruction methodology and highlight recent research to infer information about nuclear explosions from historical measurements of atmospheric fallout.