Neural-network enabled octave-spanning coherent diffraction imaging.

The article focuses on a neural-network approach to address challenges in broadband coherent diffractive imaging (CDI) under octave-spanning ultrafast laser illumination. It introduces BP-MP-UNet, a convolutional neural network based on the UNet architecture, designed to recover monochromatic diffraction patterns from broadband patterns, enabling high-fidelity image reconstruction even with spectra spanning multiple octaves. Simulations and experiments using supercontinuum lasers demonstrate that this method outperforms traditional numerical monochromatization techniques, supports both continuous and discrete spectra, and achieves rapid reconstruction suitable for real-time imaging. The approach is applicable to extreme ultraviolet and soft x-ray sources, potentially advancing attosecond microscopy by overcoming limitations of chromatic aberrations and spectral broadening inherent in ultrafast laser imaging.