Unidirectional focusing of light using reciprocal diffractive structures optimized using deep learning

We introduce a reciprocal framework that achieves unidirectional focusing of light using a diffractive architecture composed of linear and isotropic materials. Through a gradient descent-based optimization, we engineered a cascaded stack of diffractive layers with subwavelength spatial phase features designed to produce a strongly asymmetric optical response. The resulting system efficiently focuses light energy in the forward direction while significantly attenuating it in the backward direction. A key advantage is the design’s polarization-insensitivity and high performance across a broad range of incident angles and wavelengths, which underscore its versatility. Furthermore, the system exhibits notable robustness against adversarial attacks that use external wavefront manipulation, a critical feature for secure optical applications. We experimentally validated the performance of this framework in the terahertz regime, confirming its practical viability. Importantly, the concept is readily extendable across the electromagnetic spectrum by scaling the spatial features of the diffractive elements proportional to the illumination wavelength.