Determining electronic properties from <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>L</mml:mi></mml:math>-edge x-ray absorption spectra of transition metal compounds with artificial neural networks
Johann Lüder
Abstract
X-ray absorption spectroscopy at the $L$-edge probes transitions of $2p$ electrons into unoccupied $d$ states. Applied to transition metal atoms, this experimental technique can provide valuable information about the electronic structure of $d$ states. However, multiplet effects, spin-orbit coupling, and a large number of possible transitions can cause a rather involved nature of $2p$ XAS spectra, which can often complicate extracting of information directly from them. Here, artificial neural networks trained on simulated spectra of a $2p$ XAS model Hamiltonian are presented that can directly determine information about atomic properties and the electronic configuration of $d$ states from $L$-edge x-ray absorption spectra. Moreover, the adaptable nature of artificial neural networks (ANNs) allows extending their capability to obtain information about the electronic ground state and core hole lifetimes from $2p$ XAS spectra as well as to incorporate external factors, such as temperature and experimental convolution that can affect details in spectral features. The effects of noise and background contributions in spectra on the accuracy of ANNs are discussed and the method is validated on experimental spectra of transition metal compounds, including metal-organic molecules and metal oxides.