Deep learning for isolated attosecond pulse reconstruction with the all-optical method

The characterization of attosecond pulses is crucial for attosecond metrology. In this work, we investigate the isolated attosecond pulse reconstruction with the all-optical method. The results show that this method can characterize isolated attosecond pulses with a duration shorter than 50 attoseconds. Moreover, we develop a deep learning scheme to characterize isolated attosecond pulses. Through supervised learning, the deep neural network learns the mapping from the photon spectrograms to attosecond pulses. It allows complete characterization of the amplitude and phase of isolated attosecond pulses. Compared to the conventional principal component generalized projections algorithm, the reconstruction with our neural network shows superior quality and robustness to noise. Also, the reconstruction computation time is significantly reduced to a few seconds.