Highly nonlinear ultrafast fiber amplifier beyond the L-band harnessed with machine learning

Abstract

We demonstrate amplification of ultrashort pulses in the L-band (1565–1625 nm) and U-band (1625–1675 nm) using a highly nonlinear erbium-ytterbium co-doped, small-core (3.8 μm) double-clad fiber amplifier optimized with machine learning. The fiber operates in the normal-dispersion regime, enabling a strong nonlinear broadening and a characteristic spectral redshift. A programmable optical filter controlled the seed pulse’s spectral phase and shape, while a machine-learning algorithm iteratively optimized the input conditions. This work presents the first experimental demonstration of a machine-learning-optimized, gain-managed nonlinear amplifier designed to maximize emission within a user-defined spectral range while maintaining near-transform-limited compressibility. The results pave the way for the development of smart lasers with adaptive phase and spectral shaping across previously inaccessible spectral ranges.