Artifical neural network-based predictions of surface electrocoalescence of water droplets in hydrocarbon media

Abstract

Electrowetting (EW) is a widely-studied microfluidic technique to promote coalescence of droplets as well as generate droplets. However, analytical or computational modeling of the intricate phenomena associated with electrocoalescence of water droplets in hydrocarbon media is very challenging. Presently, EW-assisted surface electrocoalescence experiments are used to train machine learning (ML) algorithms. These comprise an artificial neural network (ANN), eXtreme gradient boosting (XGBoost) and polynomial regression. These models are then used to predict the influence of parameters such as applied voltage, frequency, electrode spacing, concentration and initial droplet density normalized with uncovered area ratio (δ_i/α_i), to predict nine targets: uncovered area ratio (α_f), final droplet density normalized with uncovered area (δ_f/α_f), and seven droplet density distribution (radius) bins.

ANN was the most accurate predictive tool among the three ML models with R ² of 0.89. ANN accurately predicted the droplet distribution bins for three distinct cases of good electrocoalescence, poor electrocoalescence and satellite droplet ejection (droplet generation). SHAP (Shapley Additive exPlanations) dependence plots were used to quantify the parametric influence of various parameters on each output. For δ_f/α_f, frequency and electrode spacing were the most and least influential, respectively. Interestingly, the feature influence on droplet density distributions was observed to reverse (magnitude and direction) with droplet radius. The key novelty of this study is the coupling of data from a multifunctional microfluidic device in an ANN model to accurately predict complex stochastic droplet-related phenomena (coalescence, generation). As such, the framework developed here can be utilized for other data-rich droplet-based microfluidic systems.