Bayesian Attention Training for Shielded and Masked Nuclide Representations

Abstract

A Bayesian machine learning method for automatic extraction of optimum shielding and masking representations for nuclide identification is described. Objective is to establish a minimum set of representations per nuclide that achieves improved nuclide identification results in the presence of perturbations. Considering a comprehensive spectrum data set, conditional probabilities for specific feature areas are trained. Raw spectra and transformations are considered. It leads to formalisation of “contextual occurrence” of the feature areas, like “strong shielding”, “bare” or “masked by X” - an approach prominently known as attention training in large language models. It not only discriminates the best candidates of reference spectra for the same nuclide, but also returns more semantic information on the identified nuclide. The technique is applied to different sets of real measurements. First, the best representations of Ba-133 and weapon grade plutonium (WGPu) from various data taking campaigns are extracted, covering multiple shielding variants and masking of WGPu with Ba-133. Second, spectra from Lu-177, Lu-177m, Se-75 and Ir-192 are investigated that have strong variability in age and shielding, requiring complex reference data for being successfully identified.

Authors 3

1. Marcus J. Neuer Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen,Aachen,North Rhine-Westphalia,Germany

2. Igor Garmaev Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen,Aachen,North Rhine-Westphalia,Germany

3. F. Diel

   Ametek (Germany)

   Affiliation as printed

   innoRIID / Ametek,Research and Development,Düsseldorf,North Rhine-Westphalia,Germany