Programmable Diffractive Deep Neural Networks Enabled by Integrated Rewritable Metasurfaces

A theoretical demonstration of on-chip programmable diffractive deep neural networks using rewritable Sb₂Se₃ phase-change metasurfaces, offering a compact, non-volatile solution for optical computing.

Published in Scientific Reports, this study by Sanaz Zarei introduces a programmable on-chip diffractive deep neural network built from cascaded phase-change metasurfaces. By utilizing Antimony Triselenide (Sb₂Se₃) and direct laser writing, the architecture overcomes the static nature of traditional optical networks, allowing for non-volatile reprogramming and error correction. The design is theoretically benchmarked on MNIST digit classification and pattern recognition, achieving accuracies comparable to state-of-the-art electronic systems while maintaining an ultra-compact footprint and low power consumption.

Similar efforts can be seen in A programmable diffractive deep neural network based on a digital-coding metasurface array.

Link