CircuitS2L: Circuit Dataset Augmentation via Generative Featuring and Supervised Labeling

Learning-based techniques are increasingly integrated into electronic design automation (EDA) tool to accelerate and improve the design process. Yet, their success depends on large, diverse circuit datasets, which are scarce due to strict data privacy and confidentiality in the semiconductor industry. To address this challenge, we propose CircuitS2L, a data augmentation framework that enables task-aware generative feature augmentation for both tabular and graph-structured circuit data, producing richer and more diverse representations. CircuitS2L also introduces a supervised label generation strategy that remains robust under limited data by training in an interpretable and task-consistent manner. This ensures that the generated features and labels align with representative downstream EDA tasks, enhancing model performance and generalizability. Evaluations across multiple downstream prediction tasks show that CircuitS2L-augmented data significantly improves prediction results, achieving up to 64% MAE reduction compared to real data alone in LLM-based downstream tasks.