Reinforcement Learning for Hybrid Bonding Terminal Legalization in 3D ICs

Abstract

Hybrid bonding (HB) in 3D ICs enables scaling but introduces overlap challenges from large pitch requirements. Existing legalization methods use exhaustive sliding-window scanning, resulting in significant computational inefficiency. To address this, we propose a reinforcement learning (RL) approach that adaptively selects subregions for targeted displacement optimization. The learned policy generalizes to unseen designs without fine-tuning. Experimental results on open-source and industrial benchmarks show our method fully eliminates overlaps with minimal displacement and reduced runtime compared with baselines.