Prediction of fluid-thermal-structure interaction of three equilateral-triangular circular cylinders based on XGBoost-SHAP

An interpretable machine learning study was conducted to investigate the thermodynamic characteristics as well as the hydrodynamic characteristics of three heated equilateral-triangular circular cylinders, using the eXtreme Gradient Boosting (XGBoost) algorithm and SHapley Additive exPlanation (SHAP) analysis. A comprehensive dataset was generated through numerical simulations with variable Reynolds number (Re = 60–160), spacing ratio (L/D = 1.5–10.0), and incidence angle (α = 0°–60°) at a fixed Prandtl number (Pr = 0.7). After hyperparameter optimization, the machine learning model demonstrates an excellent predictive performance. The SHAP method was adopted to interpret the model's predictions at both global and local levels, revealing the influences of Re, L/D, and α. A detailed analysis was figured out in the combination of the instantaneous vortex structures, time-mean velocity distribution, root-mean-squared flow field, time-mean pressure coefficient, and temperature distribution. The XGBoost-SHAP methodology successfully captures the flow changes induced by the variations considered parameters. Furthermore, the explanations of wake structure interaction were provided, in accompany with the classification of flow patterns.