Large Language Models (LLMs) show great potential in automated engineering design, yet their application to microfluidic devices is limited by the absence of structured, domain-specific training data. To address this limitation, this study presents a specialized dataset generation tool designed to construct training data, starting from foundational microfluidic building blocks up to complex, real-world functional device geometries.
We developed the tool as a lightweight web-based application that operates similarly to standard vector design software. We built an intuitive canvas featuring smart shapes, including specialized channels, droplet generators, and mixers, and we integrated import and export capabilities for seamless workflow implementation. We chose SVG and JSON as the primary formats because they combine scalable visual representation with structured data that LLMs can easily parse and generate. Through this interactive interface, users can digitally recreate fundamental geometries from literature or their studies. Users design the device geometry and provide prompt-targeted descriptions specifying the functional roles of components like channels, junctions, and reservoirs. The second part of the tool compiles these user-provided designs into a structured dataset directly suitable for LLM fine-tuning.
In this study, we present a standardized pipeline that converts geometric layouts into fine-tuning-ready datasets, positioning our tool as a critical enabler for generative microfluidics. Ultimately, this work establishes the foundation required to transition from manual microfluidic design to autonomous, prompt-to-chip design.