Predicting the fluid behavior of random microfluidic mixers using convolutional neural networks

Area of Science:

• Microfluidics
• Computational Fluid Dynamics
• Machine Learning

Background:

• Numerical simulation, particularly finite element analysis (FEA), is crucial for microfluidic device design but is time-consuming.
• Previous automated design methods using FEA faced limitations in speed and design-user desire matching consistency.

Purpose of the Study:

• To develop a significantly faster method for predicting microfluidic mixer performance.
• To create a more comprehensive library of microfluidic mixer designs.
• To improve the compatibility of generated designs with user requirements.

Main Methods:

• Transformed fluid mechanics simulation into an image recognition problem.
• Developed and trained a convolutional neural network (CNN) using a pre-generated library of 10,513 microfluidic mixer designs.
• Utilized the trained CNN to predict the fluid behavior of 30,757 newly generated designs.

Main Results:

• The CNN method achieved predictions in 10 seconds, over 51,600 times faster than FEA.
• The CNN-generated library expanded to 41,270 designs, covering more possibilities in fluid velocity and concentration profiles.
• Quantitative analysis confirmed improved compatibility of the CNN library with user desires.

Conclusions:

• The CNN-based approach offers a substantial speed improvement for microfluidic mixer design and performance prediction.
• This machine learning technique enables the creation of larger, more diverse, and user-centric design libraries.
• The study presents a novel application of CNNs for fluid behavior prediction in microfluidics.