Creating an AI-driven tool to automate photonic circuit design

Photonic circuits are widely used in communications, sensing, and information processing, and with the rapid advancement of these fields, new circuit designs are needed. However, designing photonic circuits is an extremely difficult task, requiring expertise in electromagnetics, electronics, materials science, semiconductors, thermal transport, and more. This design process is also largely manual, with few available tools to automate more than the most basic tasks.

Sharma, Fu, and Ansari et al. developed a tool for converting plain-text instructions into photonic circuit designs with the help of a large language model (LLM). They hope their work serves as a proof-of-concept for a new class of tools to simplify the design process and make the field of photonics more accessible.

“We were curious to see if these tools can help automate photonic circuit design and if we can make the work easier,” said author Joyce Poon.

The team’s tool is a multi-agent framework that generates circuit layout mask files in four steps. First, an LLM converts the natural language instructions into an initial schematic. A second LLM generates a detailed design based on working components. Then, an algorithmic model turns the design into a structurally valid layout and converts it into a binary file containing the circuit’s layout data and elements, and a verification algorithm simulates the circuit.

The researchers tested their framework using models from Google, OpenAI, Anthropic, NVIDIA, and DeepSeek. Gemini 2.5, o1, and Claude Opus 4 achieved the highest success rates. Although all models performed well with simple prompts, they struggled with more complex ones.

“Right now, we’re working on improving entity extraction, the very first step in the framework,” said Poon. “We notice quite often that the interpretation fails, because [LLMs] don’t fully understand the hierarchy of the circuit that that the designer wants to have implemented.”

Source: “AI agents for photonic integrated circuit design automation,” by Ankita Sharma, Yuqi Fu, Vahid Ansari, Rishabh Iyer, Fiona Kuang, Kashish Mistry, Raisa Islam Aishy, Sara Ahmad, Joaquin Matres, Dirk R. Englund, and Joyce K. S. Poon, APL Machine Learning (2025). The article can be accessed at https://doi.org/10.1063/5.0300741 .