Abstract

Design Automation of Low-cost Microfluidic Droplet Generators

Ali Lashkaripour, PhD Candidate, CIDAR Lab, Boston University, Boston University

Droplet microfluidics advantages such as precise volume control, low reagent consumption, and high throughput have made it ubiquitous in numerous fields. However, the lack of a thorough understanding of the governing physics, a large multi-dimensional design space, and multiple droplet formation regimes have prevented the introduction of a general scaling law to predict droplet size, generation rate, and formation regime. Therefore, achieving an application specific droplet size and generation rate requires several costly and time-consuming design iterations. To address this, we used low-cost micro-milling to fabricate several droplet generators and build a large dataset of experimental data points (850 points) that cover geometry, flow condition, and performance. Using this dataset, machine learning techniques were exploited to build a neural network model that predicts the performance of a microfluidic droplet generator based on the set of inputs including geometry and flow conditions. The proposed model predicts the droplet formation regime with an accuracy of 97%. Additionally, droplet diameter was predicted with mean average errors of 9.88 Hz and 17.29 Hz for dripping and jetting regimes, respectively. Finally, droplet diameter was predicted with mean average errors of 13.36 µm and 5.25 µm for dripping and jetting regimes, respectively.