[Paper Review] Machine Learning for Electronic Design Automation: A Survey

Guyue Huang, Jingbo Hu+14 more|arXiv (Cornell University)|Jan 10, 2021

VLSI and FPGA Design Techniques167 references18 citations

TL;DR

This survey provides a comprehensive review of machine learning (ML) applications in Electronic Design Automation (EDA), organizing studies by EDA flow stages. It highlights ML's role in automating design tasks—from decision-making and performance prediction to black-box optimization and end-to-end AI-driven design—demonstrating significant improvements over traditional methods in scalability and efficiency.

ABSTRACT

With the down-scaling of CMOS technology, the design complexity of very large-scale integrated (VLSI) is increasing. Although the application of machine learning (ML) techniques in electronic design automation (EDA) can trace its history back to the 90s, the recent breakthrough of ML and the increasing complexity of EDA tasks have aroused more interests in incorporating ML to solve EDA tasks. In this paper, we present a comprehensive review of existing ML for EDA studies, organized following the EDA hierarchy.

Motivation & Objective

• To systematically review recent advances in machine learning for Electronic Design Automation (EDA) over the past five years.
• To categorize ML applications in EDA along the design flow hierarchy, including high-level synthesis, logic synthesis, physical design, and verification.
• To analyze the evolution of ML techniques in EDA, from heuristic decision support to end-to-end reinforcement learning-based automation.
• To identify key challenges such as data efficiency, domain adaptation, and model trustworthiness in industrial EDA applications.
• To outline future research directions for integrating ML more deeply into EDA tools with improved reliability and automation.

Proposed method

• Organizes existing ML-EDA studies by stages in the EDA flow: high-level synthesis, logic synthesis & physical design, mask synthesis, analog design, testing/verification, and SAT solving.
• Classifies ML applications into four types: decision-making in traditional EDA, performance prediction, black-box optimization, and automated design using end-to-end learning.
• Reviews feature engineering techniques, including image-based, geometric, graph-based, and hand-crafted features, with preprocessing steps like normalization and dimensionality reduction.
• Examines domain adaptation strategies such as cross-platform performance estimation and fine-tuning pre-trained reinforcement learning agents for new design domains.
• Discusses data efficiency techniques, including semi-supervised learning and active learning, to reduce reliance on expensive simulation or synthesis labels.
• Evaluates emerging ML models like graph neural networks (GNNs), point cloud representations, and transformer-based architectures in EDA contexts.

Experimental results

Research questions

• RQ1 How has machine learning been integrated into different stages of the EDA design flow, from high-level synthesis to physical design and verification?
• RQ2 What are the key ML paradigms used in EDA—such as supervised learning, reinforcement learning, and transfer learning—and how do they differ in automation level and performance?
• RQ3 To what extent can ML models improve design quality and reduce runtime compared to traditional heuristic-based EDA tools?
• RQ4 What challenges remain in deploying ML in industrial EDA, particularly regarding data efficiency, model generalization, and trustworthiness?
• RQ5 How can transfer learning and domain adaptation techniques reduce the need for retraining ML models across different design platforms or technologies?

Key findings

• ML-based methods have demonstrated significant improvements in design space exploration, placement, and routing, reducing runtime and improving wirelength and timing metrics.
• Performance prediction models trained on historical design data can accurately estimate design quality with minimal synthesis overhead, enabling faster design iteration.
• Reinforcement learning and black-box optimization frameworks have enabled end-to-end automation of complex EDA tasks, such as placement and routing, with results competitive to or better than traditional tools.
• Transfer learning and fine-tuning of pre-trained RL agents allow rapid adaptation to new design domains with minimal retraining, reducing deployment costs.
• Despite progress, purely ML-driven EDA tools still struggle to meet industrial reliability and accuracy requirements, highlighting the need for hybrid ML-traditional approaches.
• Trust, fairness, and interpretability of ML models remain critical barriers to industrial adoption, necessitating further research in explainable and reliable AI for EDA.