b'Spectral ObservationSophisticated deep learning enhances the value of existing spectroscopic Convolutionaltechniques by returning higher-fidelity data and making feature identification Neural Network possible where it is not presently possible.T he information that can be unraveled from radiation spectra can enable research and development of novel nuclear fuels and materials. A method to accurately and repeatably analyze spectra is crucial to spectroscopists and researchers across the nuclear industry. The conventional method of identifying and quantifying spectral features from a measurement involves tallying the counts TOTAL APPROVED AMOUNT:within characteristic energy ranges. This conventional method of analysis introduces $116,000 over 1 year uncertainties because features are fit using some form of mathematical fitting to find and then quantify features, which is never perfect. These techniques can PROJECT NUMBER:also become unreliable under certain circumstances, such as unexpected count 20A1054-014 rates, elevated background, and complex spectra. Despite substantial research to PRINCIPAL INVESTIGATOR:improve these conventional techniques, the basic method to fit a spectrum to a Ryan Fronk mathematical model has not changed dramatically for decades.CO-INVESTIGATOR: The SPectral Observation Convolutional neural networK (SPOCK) project set out to Matthew Anderson, INL explore a method to analyze collected radiation spectra using advanced, scalable deep learning. INL is uniquely equipped to shift the problem of spectroscopy away from a series of mathematical curve fits to a recognition problem using an advanced method of machine learning. INLs pairing of spectroscopic expertise with the newly built and one-of-a-kind Sawtooth supercomputer graphics processing unit cluster are essential to bring this next generation of spectroscopy using machine learning to the larger spectroscopy market.This concept was developed by training Sawtooth on real-world and simulated gamma-ray spectra. SPOCK analyzed gamma-ray spectra to greater than 99% at the 90% confidence interval. Furthermore, SPOCK was trained, tested, and operated on the International Space Stations Spaceborne Computer-2 supercomputer, returning zero errors over the course of 100 training hours. This demonstrated that SPOCK could perform: (1) autonomously in far-edge, low-wattage computing situations; and (2) in hazardous radiological environments, where interference can cause errors.TALENT PIPELINE:Robyn Hutchins, student at Kansas State UniversityHeyley Gatewood, student at Georgia Institute of TechnologyGraeme Holliday, student at University of IdahoSpaceborne Computer 2 (a) installed into the International Space Station (b). SPOCK successfully replicated terrestrial training results in a high-radiation environment.48'