Multi-Task Graph Neural Network Predictions of Auger-Electron and X-ray Photoelectron Spectroscopy

Abstract

Auger-electron spectroscopy has long accompanied x-ray photoelectron spectroscopy as a second modality to resolve chemical states with overlapping core-electron binding energies. However, analyzing the Auger spectrum is challenged by its complexity and the computational expense of its simulation. Here we demonstrate that the physical connection, and thus inter-task relationship, between the generation of a core-hole and its corresponding Auger-Meitner decay enables inductive knowledge transfer through the training of a multi-task graph neural network to predict both observables from a common graph embedding. Both task losses are combined with learned weights via the uncertainty weighting procedure. Overall, the single-task and multi-task models predict calculated and experimental Auger lineshapes with good accuracy in most cases. The performance between the two task regimes is similar, with the single-task models generally having better predictions of the finer peak structures in the spectrum. The present results demonstrate that multi-task training is a promising avenue for future developments of universal x-ray spectroscopy models with learned representations that map the molecular structure to a multitude of techniques.