Abstract:

Radio frequency (RF) circuit design demands rich experience of practical know-how and extensive simulation. Complicated interactions among different building components must be considered. This becomes more challenging at higher frequency and for sophisticated circuits. In this study, we proposed a novel design automation methodology based on deep reinforcement learning (RL). For the first time, we applied RL to design a wideband non-uniform distributed RF power amplifier known for its high dimensional design challenges. Our results show that the design principles can be learned effectively and the agent can generate the optimal circuit parameters to meet the design specifications including operating frequency range (2-18GHz), output power (>37dBm), gain flatness (<4dB) and average return loss (>5.8 dB) with GaN technology. Notably, our well-trained RL agent outperforms human expert given the same design task, with 78% accuracy and offers generalizability, which is lacked in the conventional optimization approach to shorten the time-to-market.