TEPO: A Transferable EDA Prediction Optimization Method Based on Learngene Characterization

Fanshu Sun, zhou junbo, Jing Wang, Siyuan Xu, Liying Yang

Keywords: Transfer Learning, Electronic Design Automation

TL;DR: A transferable predictive optimization method based on Learngene in the field of Electronic Design Automation

Abstract: This paper introduces TEPO, a novel multi-task learning framework to optimize Electronic Design Automation (EDA) in integrated circuit (IC) design by addressing increasing complexity and the limitations of traditional independent design task approaches. TEPO systematically decomposes design knowledge into gene knowledge and class knowledge, which are referred to as Learngenes. This framework employs a dual-pathway architecture with an adaptive gating mechanism, allowing for fine-grained control over knowledge activation and enhancing computational efficiency and interpretability. In the data input section, the VIT-GNN fusion processor, which integrates Vision Transformer (ViT) features from layout images with Graph Neural Network (GNN) features from circuit topology, spatially aligning them onto a unified 256x256 grid to preserve both global visual patterns and local structural relationships. Our approach tackles four critical challenges in EDA: knowledge fragmentation, feature integration, transferability and data scarcity. The methodology involves pre-training an upstream model to extract Learngene, which is then used to initialize a downstream 12-layer Transformer model for various prediction tasks. Experiments are conducted on CircuitNet-N28, a dataset providing multi-modal features for Congestion, DRC violations, and IR-drop prediction tasks, as well as a new thermal prediction task. The transferability of learning genes not only performs well in existing categories but also shows a faster convergence speed in new task categories. The data required for its training is also less, it saves more computing costs while achieving the same performance.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 7291