Use of Neural Networks to Unfold High Energy X-Ray Spectra

Accurate characterization of x-ray spectra is challenging yet essential for correct interpretation of data from experiments in hydrodynamics, high energy density physics, and inertial confinement fusion. A practical diagnostic for this is the Filter Stack Spectrometer (FSS), which consists of a series of filter-detector pairs. The incoming spectrum is modified by each filter, which will often produce secondary electrons, and the corresponding detector records energy deposition as an intensity-based measurement. Unfolding spectra from FSS data is complicated since the FSS response is ill-conditioned, especially when considering MeV x-rays.

In previous work we demonstrated how a Neural Network (NN) can unfold spectra for synthetic data at low energies (<1MeV) [1]. The NN predictions were highly accurate for spectra of two different distributions (exponential and gaussian) for five simple FSS designs. Here we expand the model to more complex FSS, previously fielded experimentally, and spectra distributions extending up to 40 MeV. Extending the energy range greatly increases the challenge for unfolding, as photons in the MeV range all have similar attenuation coefficients in high-Z materials. To determine the efficacy of the NN approach, we compare its reconstruction error on synthetic data to a previously developed algorithm that unfolds spectra without a priori assumptions of spectral shape [2].