Optical Multimode Tensor Processor on TFLN Platform

Abstract

Optical processors have sparked extensive research interest as candidates for post-Moore era computing hardware. Leveraging its wide bandwidth, optical computing can significantly accelerate computational speed. Moreover, it can exploit multiple multiplexing dimensions, such as time and wavelength, for multichannel parallel computing. Modes, theoretically infinitely multiplexed dimensions, hold promise for significantly increasing channel capacity in optical computing. However, research in this domain remains limited at present. Here, we have designed and fabricated a parallel convolution processor based on a multimode core on a thin-film lithium niobate (TFLN) platform. Utilizing multimode multiplexing and multimode computing cores, we achieved parallel convolution computation across four channels, demonstrating the effect of image convolution. Additionally, we trained the model using the open-source Fashion-MNIST training data set, showcasing an overall accuracy of 91.5%.