Physics-Assisted Deep Reinforcement Learning-Based TSV Array Design for Signal and Power Integrity Co-Optimization

The signal integrity (SI) and power integrity (PI) co-optimization of through silicon via (TSV) arrays, which involves two goals (SI and PI) and three objects (signal, power, and ground TSVs), has always been a challenging multi-goal and multi-object optimization problem due to the huge search space. This paper proposes a novel physics-assisted deep reinforcement learning (DRL) method to tackle this problem, which allows for co-optimizing the SI/PI of TSV arrays during a DRL training process. The verification cases show that our proposed method accommodates various optimization objectives well and achieves faster convergence and better SI/PI quality than the state-of-the-art DRL approach.