Browse by author
Lookup NU author(s): Professor Thomas PenfoldORCiD, Dr Thomas PopeORCiD
This work is licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0).
Message-passing graph neural networks have emerged as a powerful framework for machinelearning in chemistry, owing to their ability to represent molecules and materials as graphs and tolearn complex structure–property relationships directly from atomic environments. By iterativelyexchanging information between neighbouring atoms, these models capture both local chemicalinteractions and longer-range structural correlations, enabling accurate prediction of a wide rangeof molecular and materials properties. Herein, we introduce an absorber-centred graph neuralnetwork framework for X-ray spectroscopy that can employ either invariant or E(3)-equivariantmessage passing, allowing the influence of geometric symmetries on predictive performanceto be systematically assessed. Atomic environments are constructed relative to the absorbingatom, providing a physically motivated representation of the local coordination environment. Tocouple structure and spectroscopy, we incorporate an energy-conditioned attention mechanism thatdynamically weights atomic contributions as a function of photon energy, enabling the model tolearn energy-dependent structure–spectrum relationships across the absorption edge. An optionalfield-aware extension further incorporates local charge and spin information, allowing electronicstructure effects to contribute directly to the learned representation. The resulting frameworkyields accurate and transferable predictions across diverse chemical environments and establishesa flexible, symmetry-aware, and energy-resolved approach to machine-learned X-ray spectroscopy.
Author(s): Penfold TJ, Pope TJ
Publication type: Article
Publication status: Published
Journal: ChemRxiv
Year: 2026
Online publication date: 18/06/2026
Acceptance date: 18/06/2026
Date deposited: 18/06/2026
Publisher: ChemRxiv
URL: https://doi.org/10.26434/chemrxiv.15004916/v1
DOI: 10.26434/chemrxiv.15004916/v1
Notes: This is a preprint and has not been peer reviewed. Data may be preliminary.
Altmetrics provided by Altmetric
