Reconstruction of the semiconductor transition dipole moment from solid-state high-harmonic generation with a machine-learning method

The transition dipole moments between different energy bands of a crystal determine the optical properties of the solid and the light-induced electron-hole dynamics. Solid-state high-harmonic generation encodes all the information about the dipole in momentum space, making it a promising approach for reconstructing the momentum-space dipole. In this work, we develop a machine-learning framework, based on a one-dimensional semiconductor Bloch equation combined with the strong-field approximation in a single-electron picture, to reconstruct semiconductor dipoles from solid-state high-harmonic spectra. As long as the assumptions of the underlying model are satisfied, the proposed approach is robust against variations in the laser parameters, band dispersion, and hidden errors in the input data. We also discuss the potential influence of the approximations employed on the reconstruction accuracy. This method provides an alternative pathway for extracting dipole information from high-harmonic spectra.