Waveguide-array optical neural network on thin-film lithium niobate platform

Abstract

Photonic parallel computing inspires important approaches to artificial intelligence, promising substantial gains in processing speed, energy efficiency, and latency. Nevertheless, the integration density and robustness of most optical computing architectures continue to pose significant challenges. Here, we report a waveguide-array optical neural network based on thin-film lithium niobate (TFLN). By virtue of the compact size of the waveguide arrays, this architecture yields a computational density of 9.48×104 TOPS/mm2. Moreover, the high-speed and energy-efficient characteristics of the TFLN platform enable a computing speed of 135.6 TOPS with an energy efficiency of 6.28 TOPS/W. Our experiments on the iris flower and handwritten digit recognition demonstrate accuracies of 100% and 89.7%, respectively. The device achieves over 90% accuracy across a 75-nm operating bandwidth in the iris flower recognition task, indicating the wavelength insensitivity and superior robustness of the waveguide array. As a compact paradigm for optical neural networks, we demonstrate the potential of waveguide arrays in performing optical computation, thereby offering solutions for the advancement of edge computing, optoelectronic integration, and intelligent automation.