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

Abstract

We present a novel denoising scheme for spectroscopy experiments employing broadband light sources and demonstrate its capabilities using transient absorption measurements with a high-harmonic source. Our scheme relies on measuring the probe spectra before and after interacting with the sample while capturing correlations between spectral components through machine learning approaches. With the present setup we achieve up to a tenfold improvement in noise suppression in XUV transient absorption spectra compared to the conventional pump on/ pump off referencing method. By utilizing strong spectral correlations in source fluctuations, the use of an artificial neural network facilitates pixel-wise noise reduction without requiring wavelength calibration of the reference spectrum. Our method can be adapted to a wide range of experiments and may be particularly advantageous for low repetition-rate systems, such as free electron lasers as well as laser-driven plasma and HHG sources. The enhanced sensitivity enables the investigation of subtle electron and lattice dynamics in the weak excitation regime, which is relevant for studying photovoltaics and photo-induced phase transitions in strongly correlated materials.