Predicting microfluidic droplet diameters in glass capillary devices using machine learning

Abstract

Using actual experimental data of single emulsion diameters generated from glass capillary microfluidic devices, we report predictions for droplet diameters in both dripping and jetting regimes using an artificial intelligence neural network. This is the first time that glass capillary microfluidic experimental data has been employed to train a neural network and predict droplet diameters for these particular devices. The neural network inputs are fluid properties and co-flow geometries of glass capillary devices. Specifically, the inputs are orifice width, capillary spacing, flow rate ratio, inner and outer capillary number, and the regime, dripping or jetting. The experimental data set consists of 800 points; 80% of the data is for training the model and 20% is for testing. Droplet sizes generated with single emulsion devices varied from 161 microns to 1085 microns in diameter, with most droplet sizes in the dripping regime. Three model algorithms – Neural Network, Linear Regression, and Support Vector Regression – were compared for droplet diameter prediction. A neural network performed best in the dripping regime with an accuracy of 91.1%. Linear regression performed the best of the three models in the jetting regime with a 92.6% accuracy. The support vector regression, with a 10-fold cross validation average score of 0.278, was outperformed by the neural network and linear regression models with 10-fold cross validation average scores of 0.623 and 0.604, respectively. Finally, we have also successfully generated a graphic user interface that predicts microfluidic droplet diameters for glass capillary devices using the neural network.