Machine learning-aided tailoring of double-emulsions within double-T microchannel

The formation of double-emulsions or core/shell microdroplets in microchannels, essential for various chemical applications, traditionally relies on costly and time-consuming laboratory methods. In this regard, computational fluid dynamics (CFD) and artificial neural network (ANN) techniques were employed. The present study developed ANN models to predict the relationship between shell thickness and double-emulsion size in a double-T microchannel, using two datasets comprising 180 experimental and CFD data points. Assessing this relationship involved analyzing various input factors, including the Capillary, Weber (case A), and Reynolds numbers (case B) of the core, shell, and continuous phases. Among twelve training algorithms and four activation functions, the Levenberg–Marquardt (LM) algorithm with sigmoidal activation functions (Tansig and Logsig), in contrast to the linear activation functions (Poslin and Purelin), achieved the highest predictive accuracy. Additionally, the predictive accuracy of ANN models was found to be significantly improved when trained using a combination of capillary and Weber numbers, as opposed to models trained only using capillary, Weber, and Reynolds numbers. The optimal neural network architectures were [10 5] neurons for case A (tansig and logsig) and [8] neurons for case B (tansig), yielding coefficients of determination (R²) of 0.99 and 0.98, respectively. These models demonstrated high precision and effective generalization, evidenced by statistical measures such as R², MSE, RMSE, AAD, %AARD, and computational time. Moreover, their ability to generalize within the training dataset further substantiates their predictive capacity.