🦅 FALCON: An ML Framework for Fully Automated Layout-Constrained Analog Circuit Design

📣 Announcements
FALCON paper has been published at The Thirty-ninth Conference on Neural Information Processing Systems (NeurIPS) 2025! 🎉

FALCON (Fully Automated Layout-Constrained analOg circuit desigN) is a unified machine learning framework for end-to-end analog and RF circuit synthesis. Given a target performance specification, FALCON selects a suitable circuit topology and infers layout-constrained parameters through a modular, gradient-based optimization pipeline powered by graph neural networks (GNNs).

Repository Includes:

• Dataset of 1 million Cadence-simulated analog/mm-wave circuits across 20 topologies and 5 circuit families.
• ML Models for performance-driven topology selection (MLP), forward performance prediction (GNN), and parameter inference.
• Graph Representation pipeline converting Cadence netlists into rich, edge-centric graph structures.
• Layout-Aware Loss integrated into inverse design via analytical models capturing parasitic and area constraints.

FALCON achieves over 99% topology selection accuracy and sub-10% relative error in performance prediction, with layout-constrained design generation in under 1 second per instance. Its generalizability to unseen topologies and tight integration of schematic and layout objectives make it a practical foundation for fully automated, real-world analog circuit design.

📖 Table of Contents

• Environment Setup
• Usage
• Citation
• Where to Ask for Help

⚙️ Environment Setup

We recommend using Conda to manage dependencies for FALCON.

📦 Install via falcon.yml

Clone the repository and create the environment:

# Clone the repository
git clone https://github.com/AsalMehradfar/FALCON.git
cd FALCON

# Create the environment from the YAML file
conda env create -f falcon.yml

# Activate the environment
conda activate falcon

🔄 Optional: Update Environment

If you make changes to the YAML or add packages later:

conda env update -f falcon.yml --prune

🚀 Usage

FALCON consists of a three-stage pipeline for analog circuit design:

• Stage 1: Topology Selection
• Stage 2: Performance Prediction
• Stage 3: Layout-Aware Parameter Inference

Each stage can be run independently once data is prepared.

---

📂 Step 1: Preprocessing (only needed if raw data is provided)

If graphs.json is not available, first run:

python scripts/process_netlists.py

If index splits are not available, run:

python scripts/generate_splits.py

🧪 Step 2: Generate Model Inputs

To generate the MLP and GNN inputs and save the required scalers:

python scripts/save_mlp_data.py
python scripts/save_gnn_data.py

🏋️ Step 3: Train Models

To train the models for each stage of the FALCON pipeline, run:

# Stage 1: Train the MLP for topology classification
python scripts/train_mlp.py

# Stage 2: Train the GNN for performance prediction
python scripts/train_gnn.py

# (Optional) Fine-tune the GNN on unseen topologies
python scripts/finetune_gnn.py

📊 Step 4: Evaluate Each Stage

To evaluate each stage of the FALCON pipeline, run:

# Stage 1 Evaluation: Topology selection accuracy
python evaluation/mlp_eval.py

# Stage 2 Evaluation: Forward performance prediction accuracy
python evaluation/gnn_forward_eval.py

# Stage 3 Evaluation: Inverse parameter inference with layout-aware optimization
python evaluation/gnn_backward_eval.py

All scripts assume data is organized under dataset/ directory as expected.

🎯 Citation

If you use FALCON in a research paper, please cite our paper:

@inproceedings{Mehradfar2025FALCON,
      title={{{FALCON}: An {ML} Framework for Fully Automated Layout-Constrained Analog Circuit Design}},
      author={Asal Mehradfar and Xuzhe Zhao and Yilun Huang and Emir Ceyani and Yankai Yang and Shihao Han and Hamidreza Aghasi and Salman Avestimehr},
      booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
      year={2025}
}

❓ Where to Ask for Help

If you have any questions, feel free to open a Discussion and ask your question. You can also email mehradfa@usc.edu (Asal Mehradfar).