An artificial neural network model for prediction of wall temperature in vertical tube under subcritical and supercritical flow

Wall temperature prediction in vertical tubes under subcritical and supercritical flow conditions is a challenging task due to involvement of highly non-linear phenomena, such as dryout and heat transfer deterioration (HTD)-induced temperature spikes. Existing literature provides a single framework that addresses each flow condition separately. This study addresses this gap by developing a unified Bayesian Regularization Backpropagation Neural Network (BR-BPNN) model that incorporates critical heat flux conditions with 34 477 and 5535 data points for subcritical and supercritical flows, respectively. The primary goal is to avoid overfitting and achieve optimal generalization across unseen data points while developing the methodology. After a hyperparameter tuning study, the optimized network was found to contain four hidden layers with 125, 100, 75, and 50 neurons. The optimized four-layer network successfully captured the challenging physical behaviors in both flow regimes. It achieved a Root Mean Square Error (RMSE) of 6.57 % and Mean Absolute Error (MAE) of 2.84 % on unseen subcritical testing data, with 96.19 % of predictions falling within a ±10 % error margin. With 97.3 % of predictions falling within the same margin, an RMSE of 4.52 % and MAE of 2.99 % were obtained for supercritical testing data. Important non-linear trends, such as spikes caused by dryout and pseudo-critical behavior, were effectively captured by the model. This model is ideal for incorporation into digital twins, real-time thermal monitoring systems, and system-level thermal-hydraulic codes utilized in nuclear reactors, heat exchangers, and power plant safety analysis. • Unified ANN model predicts wall temperatures for subcritical and supercritical flows. • K-fold cross-validation approach used for hyperparameter tuning. • Achieves ±10 % error margin for 96.19 % of subcritical and 97.3 % of supercritical data. • Captures nonlinear behaviors like dryout with good accuracy.