Reinforcement Learning-Based Design of Interconnects With Reduced Far-End Crosstalk

The far-end crosstalk (FEXT) between transmission lines affects the circuit's functionality, and it is always essential to reduce FEXT in high-speed circuits and systems. In this article, an automatic optimization method based on proximal policy optimization (PPO) reinforcement learning is proposed with the aim of suppressing FEXT through stub grid layout optimization. PPO is used to train the stochastic policy and it can effectively handle the action space. The stub grid layout is encoded into a matrix as the input of the algorithm, which iterates automatically using a reward mechanism. After optimization, the average FEXT between neighboring transmission lines can be suppressed from –28 to –41 dB. The experimental results demonstrate that the proposed method can effectively suppress FEXT and perform well under various pseudorandom binary sequence rates in the Gbps range.