Elemental Mass Quantification from Spectral X-ray Radiographs and Fluorescence using Gauss-Newton and Deep Learning Approaches

Abstract

The goal of this thesis is to explore elemental mass quantification from spectral X-ray radiographs and X-ray fluorescence. This would provide a nondestructive technique to the IAEA for international safeguards. The entire work's setup is a 160 kVp X-ray beam incident on a powder and measured with a pixelated spectral CdTe photon detector. First, the work implements a partial-volume correction to an existing numerical approach. An alternative deep learning approach is presented using CNNs to regress elemental mass. The training dataset is generated with Monte Carlo and empirical detector characterization. An unsupervised deep learning approach is also explored for the simulation-to-experiment transformation. The method is tested on both simulation and experimental data. Lastly, X-ray fluorescence from the sample is measured with a second, out-of-beam spectral photon detector. Similarly, deep learning is used to regress elemental mass. This is done both from X-ray fluorescence alone and fused with the spectral radiographic data. The work provides new technology to the IAEA and shows how simulation can be used in deep learning where experimental data is scarce.