High-sensitivity extreme-ultraviolet transient absorption spectroscopy enabled by machine learning

TL;DR

This paper presents a machine learning approach that significantly improves the sensitivity of XUV transient absorption spectroscopy by effectively reducing noise, enabling the detection of subtle electronic and lattice dynamics.

Contribution

A novel neural network-based method for noise suppression in XUV transient absorption spectroscopy that does not require wavelength calibration of the reference spectrum.

Findings

Over tenfold noise reduction compared to traditional methods

Applicable to various beam lines for studying weak excitations

Enables investigation of subtle electron and lattice dynamics

Abstract

We introduce a machine-learning-based approach to enhance the sensitivity of optical-extreme ultraviolet (XUV) transient absorption spectroscopy. A reference spectrum is used as input to a three-layer feed-forward neural network, allowing for an efficient elimination of source noise from measurement data. In pump-probe experiments using high-harmonic radiation, we show a more than tenfold improvement in noise suppression in XUV transient absorption spectra compared to conventional referencing. Utilizing strong spectral correlations in the source fluctuations, the network facilitates a pixel-wise noise reduction without the need for wavelength calibration of the reference spectrum. The presented method can be adapted to a wide range of beam lines and enables the investigation of subtle electron and lattice dynamics in the weak excitation regime, relevant for the study of photovoltaics and photoinduced phase transitions of strongly correlated materials.