Chat-Microreactor: A Large-Language-Model-Based Assistant for Designing Continuous Flow Systems

Continuous flow in microchannels play pivotal roles in advancing automated chemical synthesis. However, frequent multiphase reactions pose key challenges for designing reactors that efficiently disperse phases. Although effort has been directed toward automating development processes, the unique nature of each existing study offers limited opportunities for generalization; consequently applying this knowledge to new microreactor designs is difficult. In this study, we introduce a self-optimizing workflow that uses large language models (LLMs) to extract key microreactor-system parameters, such as density, velocity, and Capillary number, from the literature. Minimal training led to a reduction in extraction time from 24 to 16 s, with the collected data processed by ensemble learning models delivering F1 scores greater than 70% when classifying flow patterns. Through interpretability analysis, we confirmed the machine learning model's alignment with human expertise. We then applied the model to an oil-water dispersion system for controlled alumina microsphere preparation and a piperacillin synthesis system, achieving accuracy rates of 83% and 94%, respectively. This study offers synthetic chemists an LLM-based tool for the development of microreactor systems, thereby pioneering the use of machine learning in literature-sparse fields.