Interpretable machine learning for predicting and optimizing water-in-oil droplet formation in minichannel flow

Droplet formation in two-phase minichannel flows is important in many applications, but obtaining droplets with desirable characteristics such as high sphericity and uniform size remains challenging. In this study, machine learning models were developed to predict two key droplet properties, sphericity and diameter, using CFD-generated data. A total of 198 CFD simulations were used to train and evaluate five machine learning regression models: LightGBM, AdaBoost, Random Forest, XGBoost, and Support Vector Regression. Among these models, XGBoost showed the best performance, achieving coefficients of determination of 0.978 for sphericity and 0.986 for droplet diameter. To improve model interpretability, SHAP analysis was applied and showed that viscosity and density were the most influential variables affecting both outputs. The reliability of the CFD data and the model applicability domain were assessed using the leverage approach, which indicated that only 2.2% of the droplet diameter data and 1.5% of the sphericity data were outside the applicability domain. Finally, a multi-objective NSGA-II optimization was performed to identify operating conditions that yield the desired droplet characteristics. The results show that the proposed machine learning framework can accurately predict and optimize droplet formation behavior in two-phase minichannel flow systems.