Neural-network enabled octave-spanning coherent diffraction imaging

Ultrafast lasers, providing the shortest pulses worldwide, have been playing a vital role in the ultrafast imaging technology. The temporal resolution has been increasing rapidly in recent years but finally reaches its limit—the pulse width approaches photoperiods, causing significant broadening of spectral bandwidth. The state-of-the-art high harmonics generation based attosecond lasers, with pulse widths reaching ∼50 attoseconds, present octave-spanning spectra. This brings a major challenge to traditional imaging methods, as they result in unbearable chromatic aberrations. To address this challenge, we propose the neural-network approach for broadband imaging and demonstrate its effectiveness empirically by facilitating rapid coherent diffractive imaging under octave-spanning supercontinuum illumination. The proposed method remains effective when deployed with three-octave-spanning spectra, supporting both continuous and comb-like profiles as indicated by simulations. Such lensless imaging method, applicable to both extreme ultraviolet and soft x-ray sources, potentially provides an approach to attosecond imaging.