Machine learning techniques to determine elemental concentrations from raw IBA spectra

Abstract

For many decades we have run MeV protons beams together with four IBA spectra simultaneously to obtain over 35 different elemental concentrations on any given target. These include PESA for hydrogen, RBS for carbon, nitrogen and oxygen, PIXE for aluminium to lead and PIGE for light elements like fluorine, sodium and aluminium.As part of a machine learning process we have taken five years of monthly raw spectra for each of the four IBA techniques and used the R code subroutine XgBoost and the corresponding calculated elemental concentrations from iBAT analysis code to train the system. This system training included 35 different elemental species. We then used the system to predict the elemental concentrations, from just the next six months of raw IBA spectra with no other inputs. The results were excellent for all elemental concentrations above their minimum detection limits.