Predicting Performance of Microfluidic-Based Alginate Microfibers with Feature-Supplemented Deep Neural Networks

TL;DR

This paper explores deep neural networks for predicting microfluidic fiber fabrication outcomes, enhancing accuracy by incorporating physics-based parameters despite limited data.

Contribution

It introduces a method combining dataset expansion and physics-based features to improve neural network predictions in microfluidic fiber fabrication.

Findings

Physics-based parameters improve predictive accuracy.

Dataset expansion alone does not significantly enhance predictions.

The approach addresses data scarcity issues in microfluidic modeling.

Abstract

Selection of solution concentrations and flow rates for the fabrication of microfibers using a microfluidic device is a largely empirical endeavor of trial-and-error, largely due to the difficulty of modeling such a multiphysics process. Machine learning, including deep neural networks, provides the potential for allowing the determination of flow rates and solution characteristics by using past fabrication data to train and validate a model. Unfortunately, microfluidics suffers from low amounts of data, which can lead to inaccuracies and overtraining. To reduce the errors inherent with developing predictive and design models using a deep neural network, two approaches are investigated: dataset expansion using the statistical properties of available samples and model enhancement through introduction of physics-related parameters, specifically dimensionless numbers such as the Reynolds, capillary, Weber, and Peclet numbers. Results show that introduction of these parameters provides enhanced predictive capability leading to increased accuracy, while no such improvements are yet observed for design parameter selection.