Intelligent Electronic Design Automation through Machine Learning Methods

Abstract

The exponential growth in design complexity of very large-scale integrated (VLSI) circuits driven by continuous CMOS technology scaling has created significant challenges for traditional electronic design automation (EDA) methodologies. While machine learning (ML) techniques offer promising solutions to address these limitations, their practical adoption in EDA faces three critical challenges: the complexity and time-consuming nature of ML model development, limited availability of high-quality training data due to confidentiality concerns, and security vulnerabilities of ML models to adversarial attacks.This dissertation addresses these challenges through six primary contributions:1. AutoML for EDA: An automated machine learning model development framework with application to routability prediction as a case study. 2. Privacy-preserving collaborative learning: A novel federated learning (FL) framework that enables collaborative training among design companies without explicit data sharing, addressing the critical data availability challenge. 3. HFL-LA for lithography hotspot detection: A heterogeneous FL approach with local adaptation (HFL-LA) specifically targets lithography hotspot detection while handling data heterogeneity across different design companies. 4. EDALearn: A comprehensive benchmark suite for evaluating ML-based EDA approaches with an end-to-end flow from synthesis to physical implementation. 5. DRC-guided CURE: A robust defense mechanism that defends ML-based lithography hotspot detectors against adversarial attacks using DRC-guided techniques. 6. CROP: A LLM-driven circuit retrieval and optimization framework for design-aware VLSI flow parameter tuning.These contributions significantly advance the field of machine learning for EDA by establishing methodologies for automated model development, privacy-preserving collaborative learning, robust model deployment, standardized benchmarking, and intelligent design optimization, thereby paving the way for truly intelligent electronic design automation that can adapt, learn, and operate securely and reliably in industrial environments.