Text To Agentic Photonics Chip Design

Fab-leTOME is an AI silicon photonic chip design tool that converts natural language descriptions into verified GDS and OASIS layout files, built for photonic integrated circuit engineers, researchers, and students targeting foundry tape-out. The platform runs entirely in the browser — no local toolchain or EDA installation required — and covers the full ALPINE research pipeline from entity extraction (P100) through schematic generation (P300), SAX S-parameter simulation (P400), and planned foundry targeting (P500) and tape-out packaging (P600).

Design input is plain text; the system interprets the description, selects components from a 34-cell DemoPDK library covering gratings, MZIs, ring resonators, detectors, and couplers, then generates a Graphviz schematic and a GDS or OASIS file ready for KLayout design-rule checking. Bayesian optimisation is applied to circuit parameters, and an integrated SAX + JAX + gplugins simulation stack provides S-parameter results before any file leaves the platform.

Eight LLM backends are supported including GPT-4o, Claude Sonnet 4.5, and Gemini 2.5 Pro, and a human-in-the-loop co-design agent called the Fab agent allows engineers to approve or amend decisions at each stage. The agentic system is designed to learn from feedback loops and improve parameter choices and decision paths with each iteration, not just respond to a single prompt.

Live benchmarks are published for MMI splitters, MZI free spectral range, and ring resonators on the 220 nm SOI process. Documented example circuits include WDM demultiplexers, 400G transceivers, IQ modulators, MZI switches, and ring modulator banks, giving both experienced designers and learners concrete reference paths from concept to exportable layout.

Photonic integrated circuit engineers and researchers designing silicon photonic chips from natural language to verified GDS layout in a browser-based agentic platform.