Automatic Synthesis of Light Processing Functions for Programmable Photonics: Theory and Realization

TL;DR

This paper introduces an efficient automatic method for configuring programmable photonic circuits to perform various light processing functions, leveraging gradient descent and analytical scattering matrix expressions, significantly reducing computation time.

Contribution

It presents a novel gradient-based synthesis approach that automatically configures light functions on a square-mesh PPIC without spatial separation of functions, improving efficiency and flexibility.

Findings

Achieves 3x reduction in computational cost compared to previous methods.

Can realize multiple light processing functions simultaneously on the same PPIC.

Does not require dividing the mesh into separate functional subdomains.

Abstract

Linear light processing functions (e.g., routing, splitting, filtering) are key functions requiring configuration to implement on a programmable photonic integrated circuit (PPIC). In recirculating waveguide meshes (which include loop-backs), this is usually done manually. Some previous results describe explorations to perform this task automatically, but their efficiency or applicability is still limited. In this paper, we propose an efficient method that can automatically realize configurations for many light processing functions on a square-mesh PPIC. At its heart is an automatic differentiation subroutine built upon analytical expressions of scattering matrices, that enables gradient descent optimization for functional circuit synthesis. Similar to the state-of-the-art synthesis techniques, our method can realize configurations for a wide range of light processing functions, and multiple functions on the same PPIC simultaneously. However, we do not need to separate the functions spatially into different subdomains of the mesh, and the resulting optimum can have multiple functions using the same part of the mesh. Furthermore, compared to non-gradient or numerical differentiation based methods, our proposed approach achieves 3x time reduction in computational cost.