Optical diffractive layers 3D-integrated on a silicon photonic chip for high-density optical neural networks

Abstract

Optical neural networks (ONNs) have garnered significant interest for their potential to deliver ultra-high computational bandwidth and energy efficiency. Current ONN implementations mainly fall into two categories: integrated photonic neural networks (IPNNs) and free-space diffractive optical neural networks (DONNs). However, IPNNs face challenges in network scalability due to their physical footprints and accumulated losses. DONNs offer high neuron density and spatial parallelism, but possess poor reconfigurability after fabrication. In this work, we propose, to our knowledge, a novel “2D + 3D” hybrid integrated optical neural network (HIONN) architecture that monolithically integrates a 3D optical diffractive layer onto 2D photonic chips via two-photon nanolithography. An array of vertical grating couplers facilitates the optical interconnection between the planar waveguides and 3D free space. This architecture combines the potential programmability and high-speed operation of PICs with the massive parallelism and density of 3D diffractive networks, thereby overcoming the intrinsic limitations of each implementation. As a proof of concept, we experimentally demonstrate the HIONN with a neuron density of 2.5×10 5   neurons/ mm 2 through a pattern recognition task and numerically evaluate its image edge-extraction capability. The proposed HIONN provides a proof-of-concept route toward high-density hybrid integrated optical neural networks and motivates future work on complex and high-throughput optical computing tasks.