ANN-based Analog/RF IC Synthesis Featuring Reinforcement Learning-based Fine-Tuning

Abstract

In recent years, there has been a rising interest in Analog/RF intellectual property (IP) to facilitate reusable designs and streamline the design process. However, existing analog IPs often provide a one-size-fits-all solution, limiting performance options and requiring additional re-design iterations to meet specific targets. To address this challenge, automatic synthesis tools have been employed, then the effectiveness of these tools has been significantly enhanced with the integration of artificial neural networks (ANNs). Nonetheless, the inherent inaccuracies of ANN-based circuit models still necessitate fine-tuning of the solutions. This paper introduces an efficient analog/RF circuit synthesizer that uses ANN-based models and Reinforcement Learning (RL) for fine-tuning. While the ANN-powered Pseudo-Designer immediately produces solutions for the defined constraints, RL-based optimization refines the parameters if the pre-found solution fails at the validation step. Even in worst-case scenarios, the overall design process is completed in a few minutes.