<scp>Auto‐encoder</scp> neural network incorporating <scp>x‐ray</scp> fluorescence fundamental parameters with machine learning

Abstract We consider energy‐dispersive x‐ray fluorescence (XRF) applications where the fundamental parameters method is impractical such as when instrument are unavailable. For example, on a mining shovel or conveyor belt, rocks constantly moving (leading to varying angles of incidence and distances) there may be other factors not accounted for (like dust). Neural networks do require but training neural requires XRF spectra labeled with elemental composition, which often limited because its expense. develop network model that learns from data also benefits domain knowledge by learning invert forward model. The uses transition energies probabilities all elements parameterized distributions approximate parameters. evaluate baseline models rock dataset lithium mineral exploration project. Our works particularly well some low‐Z (Li, Mg, Al, K) high‐Z (Sn Pb) despite these being outside suitable range common spectrometers directly measure, likely owing ability learn correlations nonlinear relationships.