CircuitGCL: Transferable Parasitic Estimation via Graph Contrastive Learning and Label Rebalancing in AMS Circuits

Official implementation of the paper "Transferable Parasitic Estimation via Graph Contrastive Learning and Label Rebalancing in AMS Circuits".

Code Repo

https://github.com/ShenShan123/CircuitGCL.git

📑 Table of Contents

• Overview
• Key Features
• Repository Structure
• Installation
• Usage
  • Dataset Preparation
  • Model Training
  • Evaluation
• Framework Components
• Results
• Citation
• related links
• License

🔍 Overview

CircuitGCL is a novel framework for accurate parasitic capacitance prediction in analog and mixed-signal (AMS) circuits using advanced graph neural networks. Our approach enables efficient and accurate parasitic estimation across different circuit designs, significantly reducing the need for expensive SPICE simulations during the circuit design process.

Our framework supports four key parasitic estimation tasks:

• Coupling Capacitance Regression/Classification
• Ground Capacitance Regression/Classification:

✨ Key Features

CircuitGCL addresses two major challenges in parasitic estimation:

1. Enhanced Transferability - Through self-supervised graph contrastive learning (SGRL), our model can generalize knowledge across different circuit designs
2. Robust Label Handling - Using specialized loss functions and rebalancing strategies to address the severe imbalance in parasitic capacitance distributions
3. State-of-the-Art Performance - Significant improvements in parasitic prediction accuracy compared to existing methods

📁 Repository Structure

.
├── balanced_mse.py          # Implementation of balanced MSE loss functions
├── downstream_train.py      # Training script for downstream tasks
├── main.py                  # Main entry point for the training pipeline
├── model.py                 # GNN model architecture definitions
├── requirements.txt         # Python dependencies
├── sampling.py              # Graph sampling utilities
├── sgrl_models.py           # Self-supervised graph contrastive learning models
├── sgrl_train.py            # Training script for contrastive learning
├── sram_dataset.py          # Dataset loading and preprocessing
└── utils.py                 # Utility functions

💻 Installation

Prerequisites

• Python 3.10+
• CUDA-compatible GPU (recommended)

Setup Instructions

# Clone the repository
git clone https://github.com/username/CircuitGCL.git
cd CircuitGCL

# Create and activate a conda environment
conda create -n circuitgcl python=3.10
conda activate circuitgcl

# Install dependencies
pip install -r requirements.txt

🚀 Usage

Dataset Preparation

Dataset Download Instructions

The datasets used for training and testing CircuitGCL are available for download via the following links. You can use curl to directly download these files from the provided URLs.

List of Datasets

Dataset Name	Description	Download Link
SSRAM	Static Random Access Memory dataset	Download
DIGITAL_CLK_GEN	Digital Clock Generator dataset	Download
TIMING_CTRL	Timing Control dataset	Download
ARRAY_128_32	Array with dimensions 128x32 dataset	Download
ULTRA8T	Ultra 8 Transistor dataset	Download
SANDWICH-RAM	Sandwich RAM dataset	Download
SP8192W	Specialized 8192 Width dataset	Download

Using curl to Download

To download any of the datasets using curl, you can use the following command format in your terminal:

curl -O <download_link>

For example, to download the SSRAM dataset, you would run:

curl -O https://circuitgcl-sram.s3.ap-southeast-2.amazonaws.com/raw/ssram.pt

Replace <download_link> with the appropriate URL from the table above.

Dataset Statistics

Below is a summary of the statistics for each dataset as used in our experiments:

Split	Dataset	N	NE	#Links
Train.&Val.	SSRAM	87K	134K	131K
Test	DIGITAL_CLK_GEN	17K	36K	4K
	TIMING_CTRL	18K	44K	5K
	ARRAY_128_32	144K	352K	110K
	ULTRA8T	3.5M	13.4M	166K
	SANDWICH-RAM	4.3M	13.3M	154K

Note: The number of nodes (N), edges (NE), and links (#Links) are provided for reference to give an idea of the scale and complexity of each dataset.

---

Dataset Usage

Place your circuit datasets in the ./datasets/raw/ directory. The framework supports multiple SRAM circuit designs:

datasets/raw/
├── ssram.pt
├── digtime.pt
├── timing_ctrl.pt
├── array_128_32_8t.pt
├── ultar8t.pt
├── sandwich.pt
└── ...

Model Training

Basic Training

# For node classification task
python main.py --dataset ssram+digtime+timing_ctrl+array_128_32_8t --task classification --task_level node --batch_size 128

# For edge regression task with GAI loss
python main.py --dataset ssram+digtime+timing_ctrl+array_128_32_8t --task regression --task_level edge --regress_loss gai --batch_size 128

With Contrastive Learning

# Enable contrastive learning pre-training
python main.py --dataset ssram+digtime+timing_ctrl+array_128_32_8t --task classification --sgrl 1 --cl_epochs 800

Key Arguments

Argument	Description	Options
--task	Task type	classification, regression
--task_level	Prediction level	node, edge
--dataset	Names of datasets (separated by +)	e.g., ssram+digtime
--model	GNN model architecture	clustergcn, resgatedgcn, gat, gcn, sage, gine
--sgrl	Enable contrastive learning	0 (disabled), 1 (enabled)
--regress_loss	Regression loss function	mse, gai, bmc, bni, lds
--class_loss	Classification loss function	bsmCE, focal, cross_entropy
--batch_size	Training batch size	Integer (default: 128)
--cl_epochs	Contrastive learning epochs	Integer (default: 500)

🧠 Framework Components

1. Self-supervised Graph Contrastive Learning (SGRL)

Our SGRL module enables better transferability across different circuit designs by learning circuit structure representations without relying on parasitic labels. Key features include:

• Structure-aware graph augmentation strategies
• Circuit-specific positive/negative sampling techniques
• Efficient contrastive objective functions

2. Graph Neural Network Architectures

Multiple state-of-the-art GNN architectures are supported:

• Graph Convolutional Network (GCN) - For basic graph representation learning
• GraphSAGE - For efficient neighborhood sampling and aggregation
• Graph Attention Network (GAT) - For attention-based message passing
• Residual Gated GCN - For complex edge-conditioned convolutions
• Graph Isomorphism Network with Edge features (GINE) - For enhanced expressivity
• Cluster GCN - For efficient training on large circuit graphs

3. Label Rebalancing Strategies

Several advanced techniques are implemented to handle the severely imbalanced distribution of parasitic capacitance values:

• Generative-based Adaptive Importance Loss (GAI) - Our proposed approach that adaptively weights samples based on frequency distribution
• Balanced MSE Loss - Reweights the loss based on inverse sample frequency
• Balanced Negative Sampling - Improves representation learning for rare parasitic values
• Local Distribution Smoothing (LDS) - Reduces distribution bias during training
• Focal Loss - Focuses on hard-to-predict parasitic values

📊 Results

Our approach achieves superior performance compared to existing methods in both cross-circuit parasitic prediction accuracy and handling of imbalanced parasitic distributions.

Label Distribution Analysis

Performance Improvements

Comprehensive Error Comparison

Our comprehensive experiments demonstrate significant improvements over state-of-the-art methods:

• Up to 42.48% reduction in MSE
• Up to 41.08% improvement in R²
• Consistently better performance across different circuit designs

Complete Performance Comparison Table

Testset	timectrl				array				ultra				sandwich
Metric	Loss↓	MAE↓	MSE↓	R²↑	Loss↓	MAE↓	MSE↓	R²↑	Loss↓	MAE↓	MSE↓	R²↑	Loss↓	MAE↓	MSE↓	R²↑
ParaGraph	0.0153	0.0914	0.0153	0.5250	0.0115	0.0788	0.0115	0.4252	0.0175	0.0937	0.0175	0.3200	0.0223	0.1087	0.0223	0.3389
CircuitGPS	0.0105	0.0742	0.0105	0.6911	0.0108	0.0701	0.0108	0.4576	0.0158	0.0818	0.0158	0.3845	0.0225	0.1039	0.0225	0.3326
DLPL-Cap	0.0093	0.0701	0.0093	0.7056	0.0123	0.0806	0.0123	0.3853	0.0160	0.0813	0.0160	0.3704	0.0214	0.1012	0.0214	0.3622
CircuitGCL (MSE)	0.0118	0.0868	0.0118	0.6521	0.0093	0.0671	0.0093	0.5350	0.0144	0.0794	0.0144	0.4398	0.0193	0.0992	0.0193	0.4280
CircuitGCL (BMC)	71.626	0.0628	0.0088	0.7407	74.313	0.0650	0.0092	0.5418	117.73	0.0771	0.0145	0.4350	152.34	0.0938	0.0188	0.4422
CircuitGCL (GAI)	0.0091	0.0610	0.0091	0.7300	0.0089	0.0667	0.0089	0.5556	0.0145	0.0762	0.0145	0.4358	0.0187	0.0935	0.0187	0.4445
Max. Improvement	-	33.26%	42.48%	41.08%	-	19.35%	27.64%	44.20%	-	18.68%	17.71%	37.44%	-	14.98%	16.89%	33.64%

📝 Citation

If you find this work useful for your research, please consider citing:

@article{circuitgcl2025,
  title={Transferable Parasitic Estimation via Graph Contrastive Learning and Label Rebalancing in AMS Circuits},
  author={Author1 and Author2 and Author3},
  journal={Conference/Journal Name},
  year={2025},
  publisher={Publisher}
}

📝 related links

HugFace link

https://huggingface.co/papers/2507.06535

RCG link

https://github.com/stomlvforever/RCG

arxiv: 2507.06535

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.