Machine Learning-based Capability for Radioactive and Nuclear Threat Detection and Identification

Physical Sciences Inc. (PSI) proposes to develop an Advanced Learning-Enabled Threat Search (ALERTS) software to significantly enhance detection range and source classification accuracy during search for Special Nuclear Materials and other radioactive threats. The software will implement Machine Learning algorithms for automated extraction of spectral and temporal features by training on sets of gamma spectra. The features distinguishing specific source and the radiological background will be encoded within the algorithm parameters during training and applied during search to classify spectra into relevant source classes or background. The performance of the algorithms will be extensively evaluated with respect to detection sensitivity, classification accuracy and computational efficiency. The evaluation will demonstrate 95% probability of on the move detection from 100 m stand-off range against an 8 kg WGPu source using a 2x4x16 NaI(Tl) detector. The training and validation data sets spanning the relevant search CONOPS and threat configurations will be generated from Government Furnished Data, data available to PSI as well as additional empirically collected spectra. The ALERTS software will also implement communication protocols and data formats to enable integration with existing and planned detection systems, reach-back infrastructure, and user interfaces.