Numerical modeling of the hydrodynamic behavior of MgO-water nanofluid in plate-fin heat exchangers under varying operating conditions.

This research examined the ability of an artificial neural network to predict flow behavior, defined by Reynolds number and pressure drop as functions of volume fraction and temperature, at fixed flow rates of 5, 8, and 11 LPM. The proposed framework was developed to address the nonlinear, interrelated thermo-compositional effects that influence flow inertia and pressure losses, without relying on predefined empirical relationships. The model's reliability and predictive performance were evaluated through 5-fold cross-validation, monitoring of learning errors, regression diagnostics, error quantification, and sensitivity analysis. Results from cross-validation indicated stable Reynolds number predictions at low and moderate flow rates, with mean RMSE values of 6.39 at 5 LPM and 6.16 at 8 LPM. At 11 LPM, the mean RMSE rises to 13.87, highlighting the greater sensitivity and complexity of inertia-dominated flow regimes. Conversely, predictions of pressure drop remained consistently accurate across all operating conditions, achieving mean RMSE values of 2.68 mbar, 0.56 mbar, and 0.94 mbar at 5, 8, and 11 LPM, respectively. Regression analysis confirmed a strong linear relationship between predicted and reference values for both outputs, with correlation coefficients exceeding 0.99 at low and moderate flow rates and remaining above 0.96 at the highest flow rate. Further analysis of relative and absolute errors indicated high predictive accuracy, with mean relative errors below 5% for Reynolds number and below 1.3% for pressure drop across all scenarios. Sensitivity analysis consistently identified temperature as the primary source of predictive variability, with volume fraction being the secondary contributor. Pressure drop showed significantly lower sensitivity than Reynolds number, suggesting a smoother, less nonlinear relationship with the governing inputs.