Large Language Model-Enabled Synthesis of GaAs pHEMT Distributed Amplifiers

Abstract

The design of broadband RFICs, such as distributed amplifiers (DAs), requires extensive manual effort in topology selection and parameter tuning under complex design constraints. While large language models (LLMs) have recently been explored for circuit design automation, existing approaches often suffer from limited sample efficiency and insufficient topology-level reasoning. This paper presents an LLM-enabled dual-loop multi-agent framework for the automated synthesis of DAs. The proposed framework couples an outer reasoning loop for topology generation with an inner simulation-in-the-loop Bayesian Optimization (BO) loop for efficient parameter refinement. Domain knowledge is incorporated through multi-modal extraction from prior work, while a retrieval-driven refinement mechanism enables failure-aware design updates across iterations.Implemented using the WIN 70-nm GaAs pHEMT process, the framework operates with minimal human intervention and autonomously synthesizes a three-section DA that achieves improved gain-bandwidth product (GBW) and figure of merit (FoM) relative to a human-designed baseline. These results demonstrate a promising direction toward automated broadband RFIC design.