AI Agents for Photonic Integrated Circuit Design Automation

Ankita Sharma, Yuqi Fu, Vahid Ansari, Rishabh Iyer, Fiona Kuang, K. Mistry, Raisa Islam Aishy, Shafique Ahmad, J. Matres, Dirk Englund, Joyce K. S. PoonarXiv 2025 · DOI · arXiv

TL;DR

Researchers have introduced PhIDO (Photonics Intelligent Design and Optimization), the first multi-agent AI framework capable of transforming natural-language requests into functional Photonic Integrated Circuit (PIC) layouts. By orchestrating seven different "reasoning" LLMs, the system achieves up to a 91% success rate for single devices and demonstrates a viable path toward fully autonomous, end-to-end optical chip design.

What Problem Does This Paper Address?

In the world of electronics, we have robust Electronic Design Automation (EDA) tools that handle much of the heavy lifting. Photonic Integrated Circuits (PICs)—which use light instead of electricity to process information—aren't there yet.

Currently, PIC design is a grueling, manual process. Engineers must juggle multiple physical domains (electromagnetic, thermal, and electronic) while manually selecting components, drawing schematics, and running simulations. While AI has helped with specific sub-tasks like script assistance, there hasn't been a "one-stop-shop" that takes a human-readable idea (e.g., "Design a 4-channel wavelength demultiplexer") and spits out a ready-to-fabricate GDSII layout file.

Methods

The authors developed a modular, four-stage pipeline called PhIDO:

  1. Interpreter Agent: Uses LLMs to parse natural language, extract entities (like "MZI" or "phase shifter"), and search a library of prior designs to create a high-level circuit template.
  2. Designer Agent: Maps the template to a specific Process Design Kit (PDK). It selects the best-matching physical components, configures their parameters (like waveguide width), and generates a formal schematic.
  3. Layout Module: An algorithmic stage that uses the GDSFactory toolchain to handle component placement and optical routing.
  4. Circuit Verification: A final check that simulates the circuit using Scattering (S) parameters to ensure it actually performs as requested.

To keep the AI on track, the team used Pydantic-enforced schemas (ensuring the AI's output follows a strict format) and a custom Domain-Specific Language (DSL) in YAML to act as the "glue" between human language and machine code.

Datasets

The study introduces a new benchmark testbench consisting of 102 natural language design prompts. These are categorized into four complexity levels:

  • Level 1: Single components (e.g., a single modulator).
  • Level 2: Simple 1-2 component connections.
  • Level 3: 3 to 15 components (e.g., wavelength demultiplexers).
  • Level 4: Large-scale PICs with 16 to 112 components (e.g., 64-output switch trees).

The framework was tested against seven state-of-the-art reasoning models, including OpenAI’s o1, Google’s Gemini 2.5-pro, and Anthropic’s Claude Opus 4.

Key Insights and Ideas

  • The "Thinking" Advantage: Models that use "Chain-of-Thought" (CoT) reasoning—essentially "thinking" before they speak—performed significantly better at spatial reasoning and port connectivity.
  • More Tries, Better Results: The "pass@5" metric (giving the AI five attempts) significantly boosted success rates. This suggests that many AI design failures are stochastic (random) rather than systematic; the agent often knows how to do it but might trip over a formatting error on the first try.
  • The "Best-of-Breed" Trap: Surprisingly, mixing different models for different stages (e.g., using the best Interpreter with the best Designer) actually decreased performance. The "contextual drift" between different AI "personalities" caused subtle misalignments that broke the design flow.
  • Cost vs. Performance: While OpenAI's o1 is powerful, Gemini 2.5-pro emerged as a "sweet spot," offering high success rates with significantly lower token usage and cost.

Key Results

  • Single-Device Success: For simple Level 1 designs, PhIDO achieved success rates as high as 91%.
  • Mid-Range Complexity: For circuits with up to 15 components (Level 3), the top models (o1, Gemini 2.5-pro, and Claude Opus 4) achieved an end-to-end success rate of approximately 57% (at pass@5).
  • Efficiency: Gemini 2.5-pro was noted for requiring the fewest output tokens, making it the most cost-effective "engineer" in the bunch.
  • Scaling Walls: Performance dropped sharply at Level 4 (16+ components), highlighting that while AI is great at local connections, it still struggles with the global "big picture" of massive, interconnected systems.

Limitations and Future Work

The authors are transparent about the hurdles remaining:

  • The Complexity Ceiling: Current LLMs struggle with "fan-out" circuits (like a 1x64 splitter) where they lose track of component counts and port assignments.
  • Visual Blindness: The current agents are text-based. The authors suggest that multi-modal models—capable of "looking" at schematics and diagrams—will be necessary to reach human-level design accuracy.
  • Closing the Loop: The next step is "self-driving labs," where PhIDO is integrated with robotic wafer handlers and automated testing to design, fabricate, and verify chips without human intervention.

PhIDO is currently open-source, inviting the photonics community to help expand its PDKs and knowledge base.

Summary updated

Abstract

We present Photonics Intelligent Design and Optimization (PhIDO), a multi-agent framework that converts natural-language photonic integrated circuit (PIC) design requests into layout mask files. We compare 7 reasoning large language models for PhIDO using a testbench of 102 design descriptions that ranged from single devices to 112-component PICs. The success rate for single-device designs was up to 91%. For design queries with less than or equal to 15 components, o1, Gemini-2.5-pro, and Claude Opus 4 achieved the highest end-to-end pass@5 success rates of approximately 57%, with Gemini-2.5-pro requiring the fewest output tokens and lowest cost. The next steps toward autonomous PIC development include standardized knowledge representations, expanded datasets, extended verification, and robotic automation.
BibTeX
@misc{sharma2025ai,
  author = {Ankita Sharma and Yuqi Fu and Vahid Ansari and Rishabh Iyer and Fiona Kuang and K. Mistry and Raisa Islam Aishy and Shafique Ahmad and J. Matres and Dirk Englund and Joyce K. S. Poon},
  title = {{AI Agents for Photonic Integrated Circuit Design Automation}},
  year = {2025},
  eprint = {2508.14123},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.14123},
  howpublished = {arXiv},
}
Figure