Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LuxNAS: A Coherent Photonic Neural Network Powered by Neural Architecture Search

Amin Shafiee, Febin Sunny, Sudeep Pasricha, Mahdi Nikdast

TL;DR

Coherent photonic neural networks offer high-speed, low-power computation but scale poorly with many active MZIs. LuxNAS introduces $Sb_2Se_3$-based tunable directional couplers and a neural architecture search (hybrid Simulated Annealing and LBFGS-B) to realize MAC operations with zero active power and a compact footprint, by optimizing coupling coefficients and phase shifts to approximate a target unitary. The approach demonstrates 8×8 and 16×16 networks achieving final fidelity $F(T,T')=0.75$ with $RVD<1$, and enables 4×4 transformations on an iPronics processor with up to 85% footprint reduction relative to a conventional MZI-based Clements network, indicating significant advances in scalable, energy-efficient coherent photonic computing.

Abstract

We demonstrate a novel coherent photonic neural network using tunable phase-change-material-based couplers and neural architecture search. Compared to the MZI-based Clements network, our results indicate 85% reduction in the network footprint while maintaining the accuracy.

LuxNAS: A Coherent Photonic Neural Network Powered by Neural Architecture Search

TL;DR

Coherent photonic neural networks offer high-speed, low-power computation but scale poorly with many active MZIs. LuxNAS introduces -based tunable directional couplers and a neural architecture search (hybrid Simulated Annealing and LBFGS-B) to realize MAC operations with zero active power and a compact footprint, by optimizing coupling coefficients and phase shifts to approximate a target unitary. The approach demonstrates 8×8 and 16×16 networks achieving final fidelity with , and enables 4×4 transformations on an iPronics processor with up to 85% footprint reduction relative to a conventional MZI-based Clements network, indicating significant advances in scalable, energy-efficient coherent photonic computing.

Abstract

We demonstrate a novel coherent photonic neural network using tunable phase-change-material-based couplers and neural architecture search. Compared to the MZI-based Clements network, our results indicate 85% reduction in the network footprint while maintaining the accuracy.

Paper Structure

This paper contains 3 sections, 2 figures.

Table of Contents

  1. Introduction
  2. Coherent LuxNAS Network: Device Design and Neural Architecture Search
  3. Results and Discussions

Figures (2)

  • Figure 1: An 8$\times$8 (a) MZI-based Clements network and (b) LuxNAS network performing the same unitary transformation using Sb$_2$Se$_3$ loaded tunable DCs. (c) The effective index of a strip waveguide (WG1) and Sb$_2$Se$_3$-loaded waveguide (WG2). (d) Eigen-mode expansion simulation for S21 parameter as a function of the DC coupling length. (e) Field profile of the DC when Sb$_2$Se$_3$ is partially or fully crystallized and the coupling length is 25 $\mu$m.
  • Figure 2: Cost function versus the iterations for an (a) 8$\times$8 and a (b) 16$\times$16 LuxNAS network. (c) Emulated 4$\times$4 LuxNAS and Clements networks using an iPronics SmartLight processor. $n_1$ and $n_2$ denote the total number of MZIs and DCs in the Clements and LuxNAS, respectively. (d) Output optical transmission of the two emulated networks on the iPronics platform. (e) Footprint comparison between MZI-based Clements and LuxNAS networks.