Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Democratizing Electronic-Photonic AI Systems: An Open-Source AI-Infused Cross-Layer Co-Design and Design Automation Toolflow

Hongjian Zhou, Ziang Yin, Jiaqi Gu

TL;DR

The paper tackles the barrier to deploying electronic-photonic AI systems caused by cross-layer complexity and the lack of mature EPDA tooling. It proposes an open-source, AI-infused cross-layer co-design framework anchored by SimPhony, AI-accelerated Maxwell solvers (NeurOLight, PACE, PIC$^2$O-Sim), and MAPSmaps for data, training, and inverse design, enabling hardware-aware exploration. A scalable inverse-design pipeline for meta-optical neural networks is demonstrated via SP$^2$RINT, achieving up to $1825\times$ speedups over brute-force methods and producing physically realizable metasurface designs. Collectively, these contributions democratize EPDA, reduce design cycles, and enable end-to-end optimization of electronic-photonic AI systems with push-button arch-to-layout automation.

Abstract

Photonics is becoming a cornerstone technology for high-performance AI systems and scientific computing, offering unparalleled speed, parallelism, and energy efficiency. Despite this promise, the design and deployment of electronic-photonic AI systems remain highly challenging due to a steep learning curve across multiple layers, spanning device physics, circuit design, system architecture, and AI algorithms. The absence of a mature electronic-photonic design automation (EPDA) toolchain leads to long, inefficient design cycles and limits cross-disciplinary innovation and co-evolution. In this work, we present a cross-layer co-design and automation framework aimed at democratizing photonic AI system development. We begin by introducing our architecture designs for scalable photonic edge AI and Transformer inference, followed by SimPhony, an open-source modeling tool for rapid EPIC AI system evaluation and design-space exploration. We then highlight advances in AI-enabled photonic design automation, including physical AI-based Maxwell solvers, a fabrication-aware inverse design framework, and a scalable inverse training algorithm for meta-optical neural networks, enabling a scalable EPDA stack for next-generation electronic-photonic AI systems.

Democratizing Electronic-Photonic AI Systems: An Open-Source AI-Infused Cross-Layer Co-Design and Design Automation Toolflow

TL;DR

The paper tackles the barrier to deploying electronic-photonic AI systems caused by cross-layer complexity and the lack of mature EPDA tooling. It proposes an open-source, AI-infused cross-layer co-design framework anchored by SimPhony, AI-accelerated Maxwell solvers (NeurOLight, PACE, PICO-Sim), and MAPSmaps for data, training, and inverse design, enabling hardware-aware exploration. A scalable inverse-design pipeline for meta-optical neural networks is demonstrated via SPRINT, achieving up to speedups over brute-force methods and producing physically realizable metasurface designs. Collectively, these contributions democratize EPDA, reduce design cycles, and enable end-to-end optimization of electronic-photonic AI systems with push-button arch-to-layout automation.

Abstract

Photonics is becoming a cornerstone technology for high-performance AI systems and scientific computing, offering unparalleled speed, parallelism, and energy efficiency. Despite this promise, the design and deployment of electronic-photonic AI systems remain highly challenging due to a steep learning curve across multiple layers, spanning device physics, circuit design, system architecture, and AI algorithms. The absence of a mature electronic-photonic design automation (EPDA) toolchain leads to long, inefficient design cycles and limits cross-disciplinary innovation and co-evolution. In this work, we present a cross-layer co-design and automation framework aimed at democratizing photonic AI system development. We begin by introducing our architecture designs for scalable photonic edge AI and Transformer inference, followed by SimPhony, an open-source modeling tool for rapid EPIC AI system evaluation and design-space exploration. We then highlight advances in AI-enabled photonic design automation, including physical AI-based Maxwell solvers, a fabrication-aware inverse design framework, and a scalable inverse training algorithm for meta-optical neural networks, enabling a scalable EPDA stack for next-generation electronic-photonic AI systems.
Paper Structure (8 sections, 4 figures)

This paper contains 8 sections, 4 figures.

Table of Contents

  1. Introduction
  2. Cross-Layer Co-Design For Efficient Photonic AI Systems
  3. Advanced AI-Assisted Photonic Design Automation Framework
  4. Simulation for Electronic-Photonic AI Systems
  5. Physical AI-Based Maxwell Solvers
  6. AI-Infused Inverse Design Framework
  7. Scalable Inverse Training Algorithm of Meta-Optical Neural Networks
  8. Conclusion and Outlook

Figures (4)

  • Figure 1: (a) Runtime comparison of FDFD simulation and NeurOLight framework. (b) NeurOLight learns a family of parametric Maxwell PDEs for ultra-fast optical field prediction. (c) MMI field prediction across different models. The first row shows the real part of the predicted field, the second row shows the FDFD-simulated ground-truth field, and the third row shows the prediction error.
  • Figure 2: (a) Illustration of our MAPSmaps infrastructure based on three synergistic sub-modules: MAPS-Data, MAPS-Train, and MAPS-InvDes, targeting AI-assisted photonic simulation and inverse design. (b) Top: The electrical field of the inverse-designed bend predicted by NN and verified by FDFD; bottom: Optimization trajectory driven by NN-predicted gradients. Transmission efficiency calculated based on NN-predicted fields and FDFD-simulated fields is shown for comparison.
  • Figure 3: Our proposed spatially-decoupled transfer matrix probing method SP$^2$RINT ma2025sp2rint cuts the metasurface into small patches for fast simulation that reduces complexity from cubic to linear.
  • Figure 4: $|H_z|$ field comparison fora 6-layer diffraction system consisting of 128-meta-atoms metasurfaces. SP$^2$Rint captures the transfer matrix accurately and outperforms LPA, closely matching the FDFD result.