Synthesis of X-ray spectra (SOXS): Use of machine learning techniques to generate raw X-ray spectra from known elemental concentrations

Recent developments in machine learning (ML) have enabled the creation of codes that train computers to predict elemental concentrations directly from raw IBA spectra. Once trained on spectra that span the expected elements and concentrations, these ML codes can then generate a wide range of elemental concentrations from individual raw spectra—without needing any input from physics models or detector characteristics. These methods bypass the need to use fundamental physics parameters such as ionisation cross sections, stopping powers, fluorescence yields, Coster-Kronig transitions, line emission rates together with a knowledge the detector efficiency, precise geometry, and accurate X-ray line resolution for energies from 1 to 20 keV. Here, we reverse this ML approach by using it to predict raw PIXE spectra, channel by channel, from known elemental concentrations. We train ML codes with elemental data from Al to Pb and their associated raw PIXE spectra. The trained models then generate full raw PIXE spectra from any concentration set measured on thin stretched Teflon filters. This allows the spectrum simulation of a broad range of elemental concentrations from ng/cm 2 to tens of percent and assess the detectability of overlapping or interfering trace elements.