Custom deep learning enables ultrafast pulse reconstruction using Fourier transformed collinear frequency resolved optical gating (Invited Paper)

Full temporal characterization of optical pulses is critical in understanding ultrafast phenomena including electronic transitions, laser physics, and others. We demonstrate a machine learning-based reconstruction technique to recover the complex field from frequency resolved optically gated (FROG) interferometric data collected in the collinear acquisition geometry. FROG, or the collinear version CFROG data are time-consuming to acquire, so training a high-performance ML network with measured data is challenging. The use of simulated data for training requires careful dataset construction, as we show here. We find that the combination of using the Fourier transform instead of the raw data as an input and accurate noise modeling for the synthetic-data based training are required for a robust and accurate deep-learning-based quantification. The result is a significantly faster computation than traditional methods for inverting the CFROG experimental data, potentially enabling CFROG imaging.