Machine learning based unfolding of x-ray spectra from filter stack spectrometer data

Area of Science:

• Spectroscopy
• Machine Learning
• X-ray Physics

Background:

• Filter stack spectrometers measure X-ray energy deposition via photo-stimulated luminescence (PSL).
• Accurate X-ray spectra unfolding is crucial for various scientific and industrial applications.

Purpose of the Study:

• To apply neural networks for X-ray spectra unfolding using filter stack spectrometer data.
• To evaluate the accuracy, robustness, and speed of the neural network approach.

Main Methods:

• Training a neural network on synthetic X-ray data (<1 MeV) with Maxwellian and Gaussian distributions.
• Utilizing PSL measurements from five distinct filter stack spectrometer designs.
• Testing the network's performance against ground truth spectra and simulated experimental errors.

Main Results:

• Neural network predictions closely matched ground truth spectra for single distributions, with <20% difference at high energies.
• The network demonstrated robustness to experimental errors (<5%) and some ability to unfold mixed distributions.
• Unfolding rates exceeded 1 Hz, suitable for high-repetition-rate systems.

Conclusions:

• Neural networks offer a powerful and efficient tool for X-ray spectra unfolding with filter stack spectrometers.
• The developed method is accurate, robust to experimental noise, and fast enough for demanding applications.