Optimization of Microchannel Droplet Generation Parameters Using Genetic Algorithms

Microfluidic droplet technology enables the dispersion, manipulation, and mixing of immiscible liquids within microchannels. The droplet formation process within the microchannels is governed by strong, multivariate, and nonlinear coupling effects of multiphase flow. Conventional computational fluid dynamics (CFD) simulations entail substantial computational costs, making them inadequate for rapid prediction of operating parameters based on the requirements of desired droplets. Therefore, a method for the forward ultra-fast prediction and reverse operating parameter identification of microchannel droplets is proposed, which integrates machine learning-based surrogate models with genetic algorithms. Firstly, a small dataset was constructed based on the experimentally validated three-dimensional two-phase flow CFD models of a cross-focusing microchannel, which encompasses the parameters of continuous-phase flow rate, dispersed-phase flow rate, flow rate ratio, two-phase viscosity, and interfacial tension. Spearman rank correlation analysis was employed to elucidate the significance of each physical parameter for the droplet volume. Non-significant dispersed-phase viscosity parameters (p > 0.05) were then discarded, which reduced the feature dimension. Secondly, the predictive performances of the random forest, the gradient boosting trees, the support vector regression, and the multilayer perceptron (MLP) were comparatively evaluated. The results demonstrate that the multilayer perceptron neural network achieves the best performance, with a mean absolute error (MAE) as low as 0.0049 μL and a coefficient of determination (R2) reaching 0.990. Finally, the high-precision MLP model, coupled with the genetic algorithm, was utilized to achieve global adaptive reverse optimization of flow and material property parameters based on predefined target droplet volumes. Three-dimensional CFD simulations were conducted to validate the optimized parameters. The results showed that the flow pattern remained in the squeezing–dripping regime and the relative errors of the droplet volume were within the range of −5.11% to 6.53%. The study established a fast method to predict the droplet volume in a few milliseconds, and it provided the method to achieve the operating parameters based on the desired droplet volume, which greatly improves the efficiency in the intelligent control of microfluidic droplets.