Graph-Based Reinforcement Learning Approach for Multi-Power-Domain PCB PDN Shape and Stackup Synthesis

Efficient power plane and stackup optimization is critical for Printed Circuit Board (PCB) Power Delivery Networks (PDNs), particularly in multi-power-domain designs with stringent DC Resistance (DCR) specifications. This work presents a novel reinforcement learning-based framework that assigns stackup layers for each power domain and iteratively refines power plane shapes to meet design constraints while ensuring non-overlapping layouts. The approach leverages Minimum Spanning Trees (MSTs) for initializing power plane shapes. It dynamically refines them using the A* (A-Star) algorithm with weighted pathfinding, ensuring optimal connectivity and compliance with DCR requirements. Tested extensively on multi-power-domain scenarios, the algorithm demonstrates robust performance and scalability, offering an unprecedented solution to power plane and stackup optimization challenges in PCB PDN design.