Can LLMs verify PCB designs? (autocuro.com)

Recent experiments have tested the potential of large language models (LLMs) like ChatGPT and Gemini for automating PCB design verification, which is traditionally a labor-intensive process involving the analysis of over 200 parameters. By applying these models to real STM32-based boards of varying complexity, researchers aimed to determine if LLMs could effectively automate tasks such as checking component placement and identifying routing violations. While LLMs showed some success in analyzing heavily-labeled diagrams and generating basic design checklists, they struggled with core verification tasks, particularly those requiring geometric reasoning and spatial awareness, like routing analysis and trace width verification. These findings are significant for the AI/ML community, highlighting the limitations of current LLMs in engineering applications. While they can serve as useful tools for generating documentation or discussing general design principles, they fall short in performing the detailed, context-intensive tasks necessary for PCB verification. As PCB complexity increases, LLMs become increasingly ineffective, suggesting a need for more specialized AI tools that can directly interact with structured design data. The results indicate that the future of PCB design automation may rely on purpose-built solutions rather than general-purpose LLMs, paving the way for agents designed specifically for EDA platforms.