Neural network-assisted denoising in attosecond transient absorption spectroscopy

We demonstrate an approach to denoise attosecond transient absorption measurements based on a predictive neural network. Transient absorption spectroscopy measures pump-induced absorption changes referenced to pump-off spectra. Strong correlations between the near-infrared and the corresponding high-harmonic radiation forming the attosecond probe enable the prediction of reference spectra simultaneously to the acquisition of the pump-on data. This ultimately leads to a significant noise reduction to within a factor of two of the detector's shot-noise limit. This improvement corresponds to achieving typical signal sensitivities up to an order of magnitude faster compared to the conventional scheme using the same setup and parameters - vastly reducing the common multi-hour acquisitions.