Self‐Tunable Metasurface Photoelectric Hybrid Neural Network

Abstract Metasurfaces have emerged as a transformative component in optical neural networks, enabling subwavelength‐scale light manipulation for optical computing architectures. However, their fixed parameters fundamentally limit the ability of task‐adaptive training. A self‐tunable metasurface photoelectric hybrid neural network (SMPNN) is reported. In this framework, the self‐tunable metasurface consists of a liquid crystal spatial light modulator with a phase‐only modulated metasurface, combining a digital back end and an amplitude feedback neural network (AFNN) to achieve end‐to‐end online training. The loss gradient computed from the output prediction error is backpropagated through the digital network to the optical frontend, where it guides the adjustment of liquid crystal‐driven amplitude modulation in real time. SMPNN for object classification, achieving an accuracy of 99.2% for handwritten digits and 93.7% for fashion images, results that are largely comparable to those of traditional digital neural networks is used. This co‐design paradigm unifies static metasurfaces with adaptive photonic learning, enabling scalable reconfigurable optical computing and machine vision.