This study presents a quantum annealing-aided design method for ultrathin metamaterial optical diodes. The method uses a quantum annealing-enhanced active learning scheme to automatically identify optimal designs for 130 nm-thick optical diodes. The optical diode is a stratified volume diffractive film discretized into rectangular pixels, each assigned to either a metal or dielectric material. The goal is to maximize optical isolation at specific wavelengths by optimizing the material states of each pixel. The method successfully identifies optimal structures at three wavelengths (600, 800, and 1000 nm), achieving high optical isolation with a figure-of-merit (FoM) of 0.81–0.86. The design leverages a surrogate model based on supervised machine learning to map the problem into a binary optimization task compatible with quantum annealing. The optimal structures are characterized by counterintuitive geometries that support strong surface plasmon coupling in the forward direction and weak coupling in the backward direction, resulting in high forward transmissivity and low backward transmissivity. The method is validated experimentally, demonstrating the optical isolation function of the designed diodes. The study highlights the potential of quantum annealing in optimizing complex photonic structures with high-dimensional design spaces.This study presents a quantum annealing-aided design method for ultrathin metamaterial optical diodes. The method uses a quantum annealing-enhanced active learning scheme to automatically identify optimal designs for 130 nm-thick optical diodes. The optical diode is a stratified volume diffractive film discretized into rectangular pixels, each assigned to either a metal or dielectric material. The goal is to maximize optical isolation at specific wavelengths by optimizing the material states of each pixel. The method successfully identifies optimal structures at three wavelengths (600, 800, and 1000 nm), achieving high optical isolation with a figure-of-merit (FoM) of 0.81–0.86. The design leverages a surrogate model based on supervised machine learning to map the problem into a binary optimization task compatible with quantum annealing. The optimal structures are characterized by counterintuitive geometries that support strong surface plasmon coupling in the forward direction and weak coupling in the backward direction, resulting in high forward transmissivity and low backward transmissivity. The method is validated experimentally, demonstrating the optical isolation function of the designed diodes. The study highlights the potential of quantum annealing in optimizing complex photonic structures with high-dimensional design spaces.