TLDR: PhIDO is a new multi-agent AI framework that automates the design of Photonic Integrated Circuits (PICs) from natural language requests to layout mask files. It uses large language models (LLMs) across four stages: interpreting design intent, selecting and configuring components, generating layouts, and verifying circuits. Benchmarking showed high success rates for simpler designs, with Gemini 2.5-pro being cost-effective, but performance decreased with complexity. Key challenges include interpreting vague inputs and ensuring seamless context transfer between different LLMs.
Designing Photonic Integrated Circuits (PICs) has traditionally been a complex and manual process, requiring specialized expertise across multiple physical domains like electromagnetics and semiconductors. This often involves a series of steps from initial specification to layout generation and simulation, making it time-consuming and prone to errors. Current automation tools for PIC design are limited, and the potential of AI agents in this field has largely been unexplored.
A new multi-agent framework called Photonics Intelligent Design and Optimization (PhIDO) aims to change this by automating the conversion of natural language PIC design requests into ready-to-use layout mask files. Developed by researchers including Ankita Sharma, Yuqi Fu, and Joyce K. S. Poon, PhIDO represents a significant step towards autonomous PIC development. You can find the full research paper here: AI Agents for Photonic Integrated Circuit Design Automation.
How PhIDO Works: A Modular Approach
PhIDO is built with a modular architecture comprising four main components: the Interpreter, the Designer, the Layout module, and the Circuit Verification module. These modules work together to transform a natural language input into a GDSII file, which is a standard format for integrated circuit layouts.
The Interpreter acts as the first agent, taking natural language design requests and translating them into an intermediate PIC template and a high-level schematic. It identifies key design elements, functional blocks, and connectivity, much like a human designer would break down a problem. This agent uses advanced language models and a technique called retrieval-augmented generation (RAG) to draw on curated knowledge, such as circuit templates from previous designs.
Next, the Designer agent takes the PIC template and refines it into a fully specified circuit. This involves selecting specific components from a generic silicon photonics Process Design Kit (PDK), configuring their parameters (like waveguide widths or ring radii), and generating a detailed schematic with precise port connections. The Designer ensures that the circuit is structurally valid and adheres to design constraints, even incorporating self-correction steps if errors like waveguide crossings are detected.
The Layout module then generates the actual GDSII mask layout based on the detailed circuit schematic. It handles component placement and optical routing, connecting parallel waveguides without crossings. Finally, the Circuit Verification module simulates the circuit and performs design rule checks (DRC) to ensure the layout is valid and performs as expected.
Also Read:
- Bridging AI Reasoning and Engineering Principles for Automated Analog Circuit Design
- Unlocking Analog Circuit Design for AI: Introducing Image2Net
Evaluating Performance and Future Outlook
The researchers rigorously tested PhIDO using a benchmark of 102 natural language prompts, ranging from simple single-component designs to complex PICs with up to 112 components. They compared seven different large language models (LLMs) for their performance within the PhIDO framework, including OpenAI’s o1 and o3-mini, Google’s Gemini 1.5-pro and Gemini 2.5-pro, Anthropic’s Claude Opus 4, DeepSeek’s R1, and Nvidia’s Llama-3.1-Nemotron-Ultra-253b.
For simpler designs (up to 15 components), models like o1, Claude Opus 4, and Gemini 2.5-pro showed high success rates, with Gemini 2.5-pro being particularly efficient in terms of output tokens and cost. However, performance generally decreased with increasing design complexity, highlighting areas for future improvement in structural reasoning for LLMs. The study also found that generating multiple design attempts (pass@5) significantly improved success rates, suggesting that many errors are stochastic and can be mitigated through sampling.
The Interpreter stage, responsible for understanding the initial natural language input, was identified as a major source of errors, often misinterpreting hierarchical structures or miscounting components. The researchers also noted that combining different LLMs for different stages of the workflow (a “mixed-model” approach) did not improve overall performance, indicating challenges in seamless context transfer between models.
Looking ahead, the team plans to enhance PhIDO’s ability to interpret vague inputs, potentially by incorporating multi-modal models that can reason over text, mathematical expressions, and visual data. Integrating advanced simulators and optimizers, as well as expanding the photonics-specific knowledge base, are also key future directions. Ultimately, PhIDO aims to make PIC design more accessible and to pave the way for self-driving photonics laboratories, accelerating innovation in the field.
