Improving absorption spectroscopy measurements in attosecond X-ray free electron laser experiments using machine learning

Abstract: X-ray free-electron lasers (XFELs) provide ultrashort, high-brightness pulses for probing matter on femtosecond and attosecond timescales, but their stochastic shot-to-shot fluctuations make X-ray absorption spectroscopy (XAS) challenging, as both incident and transmitted spectra are required to be measured on the same shot. Conventional approaches that rely on averaged reference spectra have limited accuracy, and full-shot characterisations are impractical. Here, we demonstrate a machine learning approach to bypass these limitations. An artificial neural network (ANN) predicts the incident spectral shape from recorded diagnostics, effectively providing the equivalent of simultaneous incident spectrum measurements based on recorded machine diagnostics. Applied to XAS of Mylar at the oxygen K-edge, the method improves the accuracy in the pre-edge region (< 530.5 eV), reducing the root-mean-squared error (RMSE) from 0.0415 OD to 0.0082 OD. This approach remains highly robust with small target datasets and reduces experimental acquisition time, offering a strategy for more accurate XAS measurements in XFEL experiments where non-invasive, single-shot monitoring of the incident spectrum is not feasible.