Autonomous Droplet Microfluidic Design Framework with Large Language Models

μ-Fluidic-LLMs: A Framework for Context-Aware Tabular Data Processing in Droplet Microfluidics

Dinh-Nguyen Nguyen; Raymond Kai-YuTong; Ngoc-Duy Dinh· ACS Omega· 2025· DOI 10.1021/acsomega.5c06253

The core problem

Droplet-based microfluidic devices offer cost-effective alternatives for biological research, and machine learning models using tabular data are increasingly used to automate their design and predict performance. However, these models often neglect contextual information such as column headings and descriptions, limiting their effectiveness. This study presents μ-Fluidic-LLMs, a framework that leverages large language models (LLMs) to capture contextual information from tabular data by converting it into a linguistic format. The goal is to enhance prediction tasks in droplet microfluidics without extensive data preprocessing.

When combined with LLMs like LLAMA3.1 and DEEPSEEK-R1, deep neural networks achieve marked improvements: mean absolute error in generation rate is lowered by nearly 40%, root mean squared error in droplet diameter is reduced by around 26%, and regime classification accuracy is enhanced by over 3% compared to prior results. These improvements are consistent across the evaluated datasets, demonstrating the effectiveness of incorporating contextual information through LLMs.

Droplet-based microfluidic devices offer cost-effective alternatives for biological research, and machine learning models using tabular data are increasingly used to automate their design and predict performance. However, these models often neglect contextual information such as column headings and descriptions, limiting their effectiveness. This study presents μ-Fluidic-LLMs, a framework that leverages large language models (LLMs) to capture contextual information from tabular data by converting it into a linguistic format. The goal is to enhance prediction tasks in droplet microfluidics without extensive data preprocessing.

The μ-Fluidic-LLMs framework processes tabular data by transforming rows and columns into natural language descriptions, which are then fed into pretrained LLMs such as LLAMA3.1 and DEEPSEEK-R1. These LLMs extract features that are subsequently used by deep neural network models for prediction tasks. The framework is evaluated on publicly available droplet microfluidics datasets, focusing on three tasks: predicting generation rate, droplet diameter, and regime classification. The transformation allows the LLM to interpret column headings and associated descriptions, thereby enriching the feature set.

The μ-Fluidic-LLMs framework overcomes the limitations of traditional tabular machine learning models by leveraging the contextual understanding of LLMs. By converting tabular data into linguistic format, the framework captures column headings and descriptions that are typically ignored, leading to more accurate predictions. The results indicate that LLMs can empower deep neural networks to be highly effective with minimal data preprocessing. This study lays the foundation for broader applications of LLMs and machine learning in microfluidic design and beyond. Future work may explore integration with other data modalities and real-time design automation.

CS practitioners and researchers

Opening member content…