Reconfigurable magneto-optical diffractive neural network with enhanced optical phase modulation

We report image classification using a diffractive neural network based on the magneto-optical effect (MO-DNN). Diffractive neural networks (DNNs) offer unique advantages such as low power consumption, high-speed computing, and parallel processing. The incorporation of the MO effect into DNNs introduces reconfigurability, nonvolatility, and the potential for compact devices. However, the phase modulation achievable with MO materials is typically smaller than that achievable with liquid crystals. In this study, we focused on the 90° polarization rotation of diffracted light induced by the MO effect and found that an MO-DNN with a single hidden layer and a polarizer achieved 98% classification accuracy for the MNIST handwritten digit dataset. The MO-DNN was physically implemented using a bismuth, gallium-substituted garnet film as the MO medium and a thermomagnetic recording technique. Its performance was demonstrated experimentally with a classification accuracy of 83%, and task switching was achieved by rewriting the MO hidden layer. The MO-DNN demonstrates strong potential for realizing compact, reconfigurable, and energy-efficient photonic computing devices integrated with image sensors.