Multichannel meta-imagers based on electrically tunable metasurfaces for accelerating matrix operations

Optical neural networks have facilitated groundbreaking advancements in machine vision, intelligent healthcare, and autonomous driving systems. However, fixed optical network architectures encounter a dramatic challenge in achieving large-scale parameter tuning without necessitating modifications to the physical components, thereby limiting their capacity to handle intricate application scenarios. Here, a meta-imager based on tunable metasurfaces is proposed, which acts as an optoelectronic front end in collaboration with a digital back end to transfer computationally intensive convolution operations to super integration, highly parallel optical devices. The tunable metasurface integrated liquid crystal (LC) within our system allows for the multiplexing of both polarization and angle, generates multiple convolution channels, and performs both positive and negative convolution operations simultaneously. Moreover, two helicity channels could be tuned by voltages to exponentially double the number of information channels thus overcoming the limitations of fixed information channels in traditional metasurfaces. Our tunable meta-imager achieves 98.5% accuracy for handwritten digitals and 90.9% accuracy for fashion images. Due to its high speed, compactness, and tunable characteristics, this system holds significant potential in the field of artificial intelligence (AI).