Autonomous Droplet Microfluidic Design Framework with Large Language Models.

Area of Science:

• Microfluidics
• Machine Learning
• Bioengineering

Background:

• Droplet-based microfluidic devices offer cost-effective biological research tools.
• Machine learning models automate microfluidic device design but often miss crucial contextual data in tables.

Purpose of the Study:

• To present μ-Fluidic-LLMs, a framework for extracting contextual information from tabular data in microfluidics.
• To enhance machine learning model performance in predicting microfluidic device design and efficiency.

Main Methods:

• Transforming tabular microfluidic data into a linguistic format.
• Leveraging pretrained large language models (LLMs) for feature extraction and analysis.
• Evaluating the μ-Fluidic-LLMs framework on publicly available droplet microfluidic datasets.

Main Results:

• The μ-Fluidic-LLMs framework significantly improves deep neural network (DNN) performance with minimal data preprocessing.
• Combined with LLAMA3.1 and DEEPSEEK-R1, DNNs showed a ~40% reduction in mean absolute error for generation rate and a ~26% reduction in root mean squared error for droplet diameter.
• Regime classification accuracy improved by over 3% compared to previous methods.

Conclusions:

• μ-Fluidic-LLMs effectively captures contextual information from tabular data, overcoming limitations of traditional machine learning models.
• This framework empowers DNNs for microfluidic applications, demonstrating substantial improvements in prediction accuracy and efficiency.
• The study highlights the potential of LLMs and machine learning in advancing microfluidics research and development.