Mode Division Multiplexing (MDM) Weight Bank Design for Use in Photonic Neural Networks

TL;DR

This paper explores the design and experimental validation of mode division multiplexing (MDM) weight banks for photonic neural networks, aiming to enhance information capacity and enable scalable, real-time photonic computing.

Contribution

It introduces three experimental designs for integrating MDM into photonic neural networks, including optimal waveguide geometry, combined MDM and WDM weight banks, and a scalable neuron network topology.

Findings

Optimal waveguide geometry for mode coupling identified

Successful integration of MDM and WDM in weight banks

Prototype neuron network demonstrates scalable topology

Abstract

Neural networks provide a powerful tool for applications from classification and regression to general purpose alternative computing. Photonics have the potential to provide enormous speed benefits over electronic and software networks, allowing such networks to be used in real-time applications at radio frequencies. Mode division multiplexing (MDM) is one method to increase the total information capacity of a single on-chip waveguide and, by extension, the information density of the photonic neural network (PNN). This Independent Work consists of three experimental designs ready for fabrication, each of which investigates the process of expanding current PNN technology to include MDM. Experiment 1 determines the optimal waveguide geometry to couple optical power into different spacial modes within a single waveguide. Experiment 2 combines MDM and previous wavelength division multiplexing (WDM) technology into a single weight bank for use as the dendrite of a photonic neuron. Finally, Experiment 3 puts two full neurons in a folded bus, or "hairpin," network topology to provide a platform for training calibration schemes that can be applied to larger networks.