Title: pcbGPT: Enabling Automatic PCB Schematic Synthesis via Natural Language Inputs

Abstract

Converting natural-language hardware specifications into accurate printed circuit board (PCB) schematics continues to pose significant challenges for developers working in embedded systems, IoT, and wearable technology. Before layout and prototyping can commence, engineers are tasked with selecting compatible parts, interpreting technical datasheets, integrating support circuitry, and ensuring correct interface definitions. Furthermore, many of these circuits lack straightforward simulation methods for validation. To address this, we introduce pcbGPT, a grounded framework designed to produce editable KiCad schematics directly from natural-language requirements.

The system represents circuit designs using a Python-based domain-specific language (DSL). It integrates tool-augmented synthesis with component library searches, design knowledge grounded in datasheets, execution-based verification, and both structural and semantic validation. Additionally, pcbGPT features an interactive web interface that facilitates iterative refinement and synchronization with KiCad projects.

We assessed the system’s performance on 20 embedded schematic-generation tasks, utilizing reference implementations, specific component requirements, and interface constraints to allow for automatic comparison. The top-performing model achieved an overall pass@1 score of 0.90 and a pass@5 score of 1.00. Performance varied by difficulty: pass@1 reached 1.00 for basic and easy tasks, 0.91 for medium-level tasks, and 0.72 for hard tasks. Analysis of failures indicates that while pcbGPT is capable of generating useful, reviewable first-draft schematics for early-stage prototyping, it has not yet reached a level of reliability sufficient to entirely replace expert engineering review.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC