Inverse‐Designed On‐Chip Orbital Angular Momentum Mode Converter for Optical Convolution Acceleration

TL;DR

This paper introduces an on-chip device that efficiently converts and multiplexes orbital angular momentum (OAM) modes to accelerate optical neural networks.

Contribution

The novel contribution is an inverse-designed on-chip OAM mode converter and multiplexer with high efficiency and integration into an optical convolutional neural network.

Findings

The OAM mode converter achieves up to 88.68% up-conversion efficiency and 88.04% down-conversion efficiency.

An OAM-encoded optical CNN achieves 98.0% accuracy on MNIST and 86.1% on Fashion-MNIST classification.

Abstract

Optical neural networks leverage the inherent parallelism of light to multiplex across various degrees of freedom including wavelength, polarization, and modes. Among these, orbital angular momentum (OAM), possessing a theoretically infinite number of orthogonal mode dimensions, holds significant potential for constructing optical neural networks. However, OAM conversion and multiplexing on integrated photonic chips remain challenging. Here, we present an on‐chip OAM mode converter and multiplexer device based on inverse design. The OAM mode converter achieves maximum up‐conversion efficiency of 88.68% (OAM−1→−2), maximum down‐conversion efficiency of 88.04% (OAM−3→−1), and maximum modulation depth of 4.07 dB (OAM+1→+3). Besides, the OAM±1,±2 multiplexer achieves maximum conversion efficiency of 98.29% and maximum modulation depth of 20.69 dB. Subsequently, we demonstrate an OAM‐encoded hybrid optical convolutional neural network built using this device, achieving 98.0% accuracy on MNIST handwritten digit recognition and 86.1% accuracy on Fashion‐MNIST classification. This device provides a novel approach for on‐chip OAM conversion and multiplexing while also enabling on‐chip optical convolution operations by using OAM mode. This work offers a practical pathway for integrating OAM with on‐chip optical neural networks.

This work proposes OAM‐encoded CNN based on cascaded on‐chip OAM converters and multiplexers. The footprint of inverse‐designed converters is 5 μm × 4 μm and the maximum conversion efficiency of OAM reaches 88%. The OAM‐encoded optical CNN was built using this device, achieving 98.0% accuracy on MNIST and 86.1% accuracy on Fashion‐MNIST classification. This device offers a practical pathway for integrating OAM with on‐chip optical neural networks.

Figures