Engineering spectro-temporal light states with physics-trained deep learning

Frequency synthesis and spectro-temporal control of optical wave packets are central to ultrafast science, with supercontinuum (SC) generation standing as one remarkable example. Through passive manipulation, femtosecond (fs) pulses from nJ-level lasers can be transformed into octave-spanning spectra, supporting few-cycle pulse outputs when coupled with external pulse compressors. While strategies such as machine learning have been applied to control the SC's central wavelength and bandwidth, their success has been limited by the nonlinearities and strong sensitivity to measurement noise. Here, we propose and demonstrate how a physics-trained convolutional neural network (P-CNN) can circumvent such challenges, showing few-fold speedups over the direct approaches. We highlight three key advancements enabled by the P-CNN approach: (i) on-demand control over spectral features of SC, (ii) direct generation of sub-3-cycle pulses from the highly nonlinear fiber, and (iii) the production of high-order solitons, capturing distinct "breather" dynamics in both spectral and temporal domains. This approach heralds a new era of arbitrary spectro-temporal state engineering, with transformative implications for ultrafast and quantum science.