Firefly Demo — Interactive OPC Inference

Interactive Streamlit app for Project Firefly V4. Draw semiconductor geometry, simulate hardware degradation, and run the physics-informed U-Net to generate a corrected binary mask.

Launch →
Hosted on Streamlit Community Cloud — may take ~30 seconds to wake up if inactive.

---

What it does

The app simulates three stages of the lithography pipeline:

1. INPUT — Damaged Wafer: Your drawn mask with random emitter failures applied (pixel dropout).
2. CORE — ILT Prediction: The U-Net reconstructs a binary OPC mask, placing serifs and dog-ears at corners to pre-compensate for diffraction.
3. OUTPUT — Printed Silicon: A Gaussian blur over the corrected mask simulates the final aerial image on the wafer.

---

How to use it

1. Draw geometry in the DESIGN_VIEWPORT using Line or Block tools. White = material deposited, Black = etched away (Manhattan geometry only).
2. Set physics parameters in the sidebar:
   • Emitter Failure Rate (%) — percentage of emitter pixels randomly killed to simulate hardware degradation.
   • Diffraction Blur (Sigma) — controls the width of the simulated optical PSF.
3. Click EXECUTE NEURAL OPC to run inference.
4. Read the three-panel TELEMETRY_READOUT output.

---

Model

Property	Value
Architecture	U-Net (3-level encoder-decoder)
Input	2-channel: damaged mask + original mask (256×256)
Output	Binary chrome mask (256×256)
Parameters	~384K
Training data	Synthetic Manhattan geometry with simulated stochastic emitter failure
Inference time	<50ms on CPU

The model file is at models/saved/firefly_brain_v4_fourier.pth.

For the full technical description of the optical model and training setup, see the whitepaper.

---

Run locally

git clone https://github.com/BotickDev/Firefly-Demo.git
cd Firefly-Demo
pip install -r requirements.txt
streamlit run streamlit_app.py

Requirements: Python 3.9+, PyTorch (CPU is sufficient), see requirements.txt for full list.

---

Repository structure

Firefly-Demo/
├── models/
│   └── saved/
│       └── firefly_brain_v4_fourier.pth   # Trained U-Net weights
├── streamlit_app.py                        # App entrypoint
└── requirements.txt

---

Related

• FireFly-Showcase — results, documentation, and whitepaper.