ConvNet-Based Prediction of Droplet Collision Dynamics in Microchannels

TL;DR

This paper introduces a CNN-based method to predict droplet collision outcomes in microchannels, significantly improving speed and accuracy over traditional simulation techniques.

Contribution

The study presents a novel application of convolutional neural networks to predict droplet collision results, reducing computational time and enhancing scalability.

Findings

Prediction accuracy of 97.2% on test data

CNN model outperforms traditional methods in speed and scalability

Effective in varied physical parameter conditions

Abstract

The dynamics of droplet collisions in microchannels are inherently complex, governed by multiple interdependent physical and geometric factors. Understanding and predicting the outcomes of these collisions-whether coalescence, reverse-back, or pass-over-pose significant challenges, particularly due to the deformability of droplets and the influence of key parameters such as viscosity ratios, density ratios, confinement, and initial offset of droplets. Traditional methods for analyzing these collisions, including computational simulations and experimental techniques, are time-consuming and resource-intensive, limiting their scalability for real-time applications. In this work, we explore a novel data-driven approach to predict droplet collision outcomes using convolutional neural networks (CNNs). The CNN-based approach presents a significant advantage over traditional methods, offering faster, scalable solutions for analyzing large datasets with varying physical parameters. Using a lattice Boltzmann method based on Cahn-Hilliard diffuse interface theory for binary immiscible fluids, we numerically generated droplet collision data under confined shear flow. This data, represented as droplet shapes, serves as input to the CNN model, which automatically learns hierarchical features from the images, allowing for accurate and efficient collision outcome predictions based on deformation and orientation. The model achieves a prediction accuracy of 0.972, even on test datasets with varied density and viscosity ratios not included in training. Our findings suggest that the CNN-based models offer improved accuracy in predicting collision outcomes while drastically reducing computational and time constraints. This work highlights the potential of machine learning to advance droplet dynamics studies, providing a valuable tool for researchers in fluid dynamics and soft matter.