Deep learning reconstruction of attosecond X-ray pulses from an angularly streaked 2D photoelectron momentum distribution

We present a deep neural network (NN) to reconstruct attosecond X-ray pulses using the photoelectron momentum spectra (PEMS) from a two-color (X-ray/IR) field. A circularly polarized IR field maps the temporal profile of the X-ray pulse onto the PEMS, which is projected onto a 2D detector using a coaxial velocity map imaging spectrometer (cVMI). Our NN uses the 2D PEMS to predict the electric field of the X-ray pulse. We trained the 5-layer, fully connected network on simulated cVMI data, and tested the NN on experimental cVMI data taken at the Linac Coherent Light Source to benchmark against existing techniques. NN reconstruction of attosecond pulses from such cVMI projections allows for fast characterization of pulses, with possible application to real-time pulse diagnosis at XFELs.