A data driven framework for optimizing droplet microfluidics with residual block and Fourier enhanced networks.

Area of Science:

• Microfluidics
• Biomedical Engineering
• Computational Science

Background:

• Droplet-based microfluidic devices are crucial for diagnostics and pharmaceutical screening.
• Precise control over droplet size and production rate is essential for efficiency and cost reduction.
• Traditional microfluidic device design methods are time-consuming and expensive due to reliance on complex models or trial-and-error.

Purpose of the Study:

• To develop a data-driven framework for predicting droplet characteristics and optimizing microfluidic device geometry.
• To integrate forward prediction of droplet size and frequency with inverse design for geometric optimization.
• To reduce the time and cost associated with microfluidic device design and prototyping.

Main Methods:

• Simulated droplet formation dynamics in a co-flow microfluidic device using the Lattice Boltzmann method, generating 658 cases.
• Developed two machine learning models: Fourier-Enhanced Network (FEN) and Residual Block Network (ResBNet).
• FEN utilizes Fourier series for feature decomposition; ResBNet uses residual blocks with skip connections for complex patterns.

Main Results:

• ResBNet accurately predicted relative droplet radius, Strouhal number, and optimized geometric ratios.
• FEN demonstrated computational efficiency and robust performance across various flow rates.
• The developed DesignFlow platform automates microfluidic device design and optimization.

Conclusions:

• The data-driven framework significantly reduces simulation time and cost for microfluidic device design.
• The developed machine learning models and DesignFlow platform enable rapid prototyping for biomedical and pharmaceutical applications.
• This approach offers a more efficient and cost-effective alternative to traditional microfluidic device design methodologies.