Content area
Abstract
Printed Circuit Boards (PCBs) serve as the foundation of electronic products, facilitating the physical and electrical integration of electronic components. PCB routing, a crucial step of design, computationally determines optimal paths for metal traces based on component placement and connectivity requirements. However, with advancing integrated circuit technology, the complexity of routing increases, leading to time-consuming and error-prone processes. Current PCB routing paradigms often separate routing into escape and area routing stages, but this isolation can result in suboptimal solutions. In addition, existing routing algorithms are typically specialized and lack adaptability to address evolving design needs. Moreover, the absence of standardized benchmarks impedes the assessment of new routing approaches, particularly those incorporating machine learning techniques. To tackle these challenges, this dissertation proposes novel routing algorithms and a comprehensive dataset of real-world PCB designs.
The research unfolds in three distinct parts. In the first part, we propose an end-to-end solution using Monte Carlo tree search (MCTS) and deep reinforcement learning (DRL). This approach employs a designed MCTS for circuit routing, guided by a deep reinforcement learning policy in the rollout. Notably, our method can be easily extended to routing cases with diverse routing constraints and optimization goals. Experimental results underscore the superiority of our approach, showcasing the highest success rates and minimized total wirelengths compared to traditional sequential A*-based routers on the test set.
In the second part, we introduce a pad-focused, net-by-net, two-stage PCB routing methodology. This comprises an MCTS-based global routing stage followed by an A*-based detailed routing stage. To minimize the gap between the proposed global and detailed routing, a polygon-based dynamic routable region partitioning mechanism is introduced to guarantee the existence of a detailed routing solution when a global routing solution exists. Our experiments demonstrate the outperformance of our approach against both state-of-the-art academic and non-academic routers, evident in terms of enhanced routability and reduced wirelength.