MicrowellMicrofluidicsMiner (M3): Leverage Large Language Model Agents for Knowledge Mining of Microwell Microfluidics

Microwell microfluidics has emerged as powerful platforms for high-precision biological and chemical investigations, bridging microscale fluid handling with compartmentalized reaction environments. Achieving robust and reproducible performance in such studies requires substantial effort to optimize microwell array design. This burden could be markedly alleviated by the availability of a curated database of microwell array parameters. Such a resource would enable the application of machine-learning models for performance prediction and automated design, leveraging knowledge accumulated from prior microfluidics research. However, constructing such a database entails a considerable investment of time and extensive manual curation, as microwell performance is governed by numerous critical design parameters that are reported inconsistently across a broad and largely unstructured body of literature. In this study, we introduce MicrowellMicrofluidicsMiner (M3), a framework that employs large language model (LLM) agents for autonomous knowledge extraction in microwell microfluidics. To evaluate its performance, we curate a ground-truth database and establish an LLM-driven assessment approach. Our results demonstrate that M3 achieves a peak accuracy of approximately 78%, representing more than a twofold improvement over the lowest observed accuracy (32%) obtained using a standalone LLM model (LLAMA 3.1). This study provides a foundational reference for researchers seeking to apply LLM agents to data-driven microfluidics research. The insights presented have the potential to substantially improve how scientists across microfluidics-related disciplines access, interpret, and leverage scientific information, thereby accelerating the development of innovative microfluidic devices and associated discoveries.