Machine learning approach in predicting nanofluid viscosity of alumina, copper oxide, silicon dioxide, and titanium dioxide using physics constraint-based XGBoost model

Nanofluids are essential colloidal suspensions composed of base fluids with suspended nanoparticles. They possess enhanced thermophysical properties, making them useful for various applications, including heat exchangers, solar collectors, and cooling systems. Although viscosity is a key property that affects heat transfer in nanofluids, it is difficult to predict accurately. Experimental methods provide precise viscosity values but are costly and time-consuming. Therefore, machine learning models have been developed to predict the viscosity of nanofluids more efficiently. The purpose of this research is to develop a Physics-Guided Extreme Gradient Boosting (XGBoost) model (PGXGB) for estimating nanofluid viscosity by incorporating physics-based relationships into the conventional loss functions of XGBoost algorithm. The model is applied to predict the viscosity of water-based Al2O3, TiO2, SiO2, and CuO nanofluids using 792 experimental data points. The results are compared with other machine learning models such as Gradient Boosting Regressor, Extra Tree Regressor, and Nu-Support Vector Regressor. The proposed PGXGB model accurately predicts all data points, demonstrating excellent accuracy and very low prediction error (R2 = 0.992 248, RMSE = 0.0559 336). Furthermore, statistical and graphical error evaluations demonstrate that the PGXGB model outperforms widely cited empirical, theoretical, and soft computing models in terms of both accuracy and validity range. Sensitivity analysis is also conducted to determine the input factors that most significantly affect prediction accuracy. Among the individual variables, nanoparticle volume fraction is found to have the greatest influence.