Machine learning-based optimization and shap analysis of material, pressure, and fluid type effects on different outlet temperatures in a Ranque-Hilsch vortex tube

Abstract

Numerous studies have explored the complex dynamics and practical applications of vortex tubes (VT), with experimental approaches offering advantages in empirical validation and real-world data acquisition. In this study, a novel experimental setup was developed by connecting two counterflow vortex tubes in parallel to investigate the effects of key input parameters inlet pressure, nozzle type, and material on the hot outlet temperature (Th), cold outlet temperature (Tc), and temperature difference (Delta T). Unlike conventional methods, separate datasets were created for each output variable. Several machine learning regression models linear regression (LR), random forest (RF), K-nearest neighbors (KNNs), XGBoost (XGB), CatBoost (CatB), and gradient boosting (GB) were trained and compared to identify the most accurate predictors. SHAP (SHapley Additive exPlanations) was employed to interpret feature importance, and grid search optimization was used to fine-tune model parameters. Results showed that XGB achieved the highest accuracy (97%) for Tc, while CatB provided the best predictions for Th and Delta T, also at 97% accuracy. These findings demonstrate the efficacy of machine learning in modeling VT behavior and highlight the potential of the proposed setup for enhanced thermal performance analysis.