Deep learning-empowered metasurface design for complex target ISAR illusions

Scattering signature manipulation enabling advanced radar camouflage remains constrained in practical applications by narrow operational bandwidths, incidence-angle sensitivity, and complex analytical phase computations for geometric targets. To address these limitations, we report an inverse-design framework for the broadband inverse synthetic aperture radar (ISAR) image reconstruction of complex targets with different scattering characteristics, bypassing analytical phase derivations. The wideband phase distribution is extracted directly from the target's near-field electric field, and a cascaded forward-backward neural network is subsequently deployed to synthesize the corresponding meta-atom and array profiles. Employing the optimized metasurface, we achieve high-fidelity ISAR image illusions with a range resolution of 37.5 mm and a cross-range resolution of 15 mm across the 8-12 GHz band (X-band). The framework is validated through the broadband ISAR reconstruction of both a cross-section-invariant split metal plate and a cross-section-variant tower model, offering a robust approach for generating electromagnetic illusions of complex targets.