Utilizing large language models for automated extraction of radio-frequency circuit metrics from scientific publications

This study presents a novel methodology for constructing a comprehensive radio frequency (RF) circuit database using multimodal large language models (LLMs). The primary goal is to extract textual, graphical, and tabular data from RF-related research papers, employing advanced prompt engineering techniques to analyze and summarize key performance metrics. The approach successfully organizes data entries covering power amplifiers (PAs), low-noise amplifiers (LNAs), distributed amplifiers (DAs), filters, and switches. Additionally, this study establishes a novel scoring system to conduct a comparative analysis of the extraction capabilities of various multimodal LLMs. Furthermore, a case study on Ka-band LNA design demonstrates that the proposed global retrieval significantly expands the design space compared to manual baselines. By integrating Pareto frontier analysis with constraint verification, the framework provides rigorous insights for optimized specification definition and multi-objective trade-off assessment. • LLM-driven extraction of circuit performance metrics: We design and implement an LLM-driven methodology that automatically extracts key RF circuit parameters from academic papers. Comprehensive experiments across multiple LLMs confirm its superior accuracy and efficiency. • Holistic LLM evaluation framework: We introduce a novel, multi-dimensional scoring framework that uniquely combines rigorous quantitative metrics with a qualitative LLM-as-a-Judge methodology. This approach offers a more robust and holistic assessment than traditional accuracy metrics. • Comprehensive RF circuit database construction: We develop a large-scale, structured database of RF circuit indicators, addressing the scarcity of accessible datasets for advanced RFIC research and supporting innovation in RF design. • Data-driven design space exploration: We conduct a comprehensive case study on Ka-band LNA design. By quantitatively contrasting manual selection with LLM global retrieval, we demonstrate how the framework redefines design specifications. Furthermore, employing Pareto frontier analysis and constraint verification enables rigorous multi-objective trade-off assessment and feasibility validation.