Guiding blind signal separation in X-ray absorption spectroscopy using supervised machine learning

Transformations under working conditions experienced by heterogeneous catalysts and other functional materials commonly result in complex, time-dependent atomistic structures, where the active species that are relevant for the functionality of the material coexist with passive spectators. Decoupling the contributions of multiple coexisting species to spectroscopic data remains a common challenge for many sample-averaging techniques, including X-ray absorption spectroscopy. Here we demonstrate that a data-driven, supervised machine learning based approach can greatly assist in the challenging task of singling-out the contributions of spectroscopically distinct species. This is accomplished by recognizing common patterns in large datasets of X-ray absorption spectra that are calculated ab initio for a broad range of different compounds. Furthermore, our approach also serves as a convenient tool for the quantification of the uncertainties of spectral decomposition. We demonstrate the application of our method to model datasets and also to experimental spectra from time-resolved operando X-ray absorption measurements. The latter are carried out for a copper oxide-derived catalyst during electrocatalytic CO₂ conversion in an oxalic acid-based electrolyte. Our method reveals an intriguing charge disproportionation process, resulting in the formation of copper oxalate species and metallic Cu already under open circuit conditions.