Frequency synthesis and spectro-temporal control of optical wavepackets are central to ultrafast science

with supercontinuum (SC) generation standing as one remarkable example. Through passive manipulation

femtosecond pulses from nanojoule-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-embedded convolutional neural network that embeds spectro-temporal correlations can circumvent such challenges

resulting in faster convergence and reduced noise sensitivity. This innovative approach enables on-demand control over spectro-temporal features of SC

achieving few-cycle pulse shaping without external compressors. This approach heralds a new era of arbitrary spectro-temporal light state engineering

with implications for ultrafast photonics

photonic neuromorphic computation

and artificial intelligence-driven optical systems.