Marker Line Identification in Special Nuclear Material using Physics-Informed Autoencoding

Abstract

Nuclide identification codes have been extensively tested for their performance regarding special nuclear material (SNM) in the past. A main focus was put on the identification quality for masking scenarios. Normally strong medical masks are mixed with weak traces of the SNM, to find the limiting ratio, at which the identification code can no longer find the SNM. In the present work this perspective is changed. The paper questions, whether algorithms are capable to identify marker isotopes in SNM. Markers are of interest, as they contain information about the nuclear facility that produced the material. They consequently allow to determine a country of origin for the threat material.For the algorithm this means, the SNM is present with a strong signature. It is covering different perturbing isotopes. Advantage of this situation is, that the SNM is already known prior to the identification. A solution based on an Autoencoder (AE) is presented. This is an autoregressive neural network, that belongs to the group of unsupervised machine learning technique. Autoencoders have proven successful in anomaly detection problems in the past, as they feature two unique modes for detecting deviations from trained data.An implementation of the AE in Tensorflow is demonstrated, that is trained with over 400 real-world spectra of high enriched uranium (HEU). This network learns the best nonlinear representation of HEU in the lower dimensional subspace of the latent layers within the AE. Deviations from pure HEU can be detected quickly and the residual parts of the markers can be extracted via the AE procedure. Results of those tests are shown with a marker substitute (Eu-152). The method was also tested on Weapon-Grade Plutonium and with real marker lines.

Authors 1

1. Marcus J. Neuer Aachen Information and Automation Technology

   RWTH Aachen University

   Affiliation as printed

   InnoRIID GmbH,Research & Development,Grevenbroich,North Rhine-Westphalia,Germany

   Information and Automation Technology, RWTH Aachen, Aachen, North Rhine-Westphalia, Germany

   Research & Development, InnoRIID GmbH, Grevenbroich, North Rhine-Westphalia, Germany