Inverting the 2p XAS model Hamiltonian by artificial neural networks and application to L-edge spectra of strongly correlated systems

Johann Lueder^1*

¹Department of Materials and Optoelectronic Science, National Sun Yat-Sen University, Kaohsiung, Taiwan

* Presenter:Johann Lueder, email:johann.lueder@gmail.com

X-ray absorption spectroscopy (XAS) is an element- and state-selective technique, which can reveal detailed information on the nature of d-electrons when applied to the L-edge of transition metals. For this excitation process, a large number of possible transitions obeying dipole selection rules, multiplets, spin-orbit coupling (SOC), temperature and lifetime effects, as well as the dependence on the chemical surrounding of the metal ion, can result in complicated and feature-rich spectra. Often, a manual trail-and-error fitting procedure reproducing an experimental spectrum with a model Hamiltonian is used to elucidate electronic states and effects. This can be labour intensive and time-consuming because the employed model Hamiltonian can have, in some cases, ten or more parameters that must be explored simultaneously.
Here, an artificial neural network-based approach is presented that can reliably determine a set of most suitable parameters describing the electronic state of the metal ion and its surrounding from an experimental spectrum. Then, these parameters together with a 2p XAS model Hamiltonian can reproduce the given spectrum, reaching a very high degree of similarity [1]. Moreover, the effect of background signals and noise is discussed as well as how to include temperature and experimental broadening conditions into the method.

(1) Lueder, J. Determining Electronic Properties from L-Edge X-Ray Absorption Spectra of Transition Metal Compounds with Artificial Neural Networks. 2020. ArXiv:2009.09684

Keywords: X-ray absorption spectroscopy, Machine learning, transition metals