Prediction of pressure drop in hexagonal wire-wrapped rod bundles using artificial neural network

Abstract

In this study, an artificial neural network (ANN) model was utilized to predict the friction factor for wire-wrapped rod assemblies. The 80 experimental data sets from UCTD correlations (Chen et al., 2018) were used to train and validate the ANN models. Three ANN models were introduced for laminar, transition, and turbulence flow conditions, respectively. The flow regime was determined based on the Reynolds number and the pitch to diameter ratio. To estimate the friction factor, Reynolds number, and multiple design parameters such as the number of rods, the rod's diameter, the diameter of the wire, the lattice pitch, the edge pitch, and the wire's helical pitch were used from the 80 bundles dataset. Three-quarters of the total data was used for training, while the other quarter was used for validation. The Levenberg-Marquardt method with Gauss-Newton approximation for Hessian of the training cost function was applied for training. For validation of the models, the cross-validation method was adopted. The ANN models were composed of seven inputs, R1 neurons in the hidden layer, and a single-output (7-R1-1). The R1 was determined based on the minimum validation error principle. Training and validation were repeated 20 times using randomly shuffled data sets to estimate the ANN model's uncertainty and errors. The prediction of the trained models showed meaningfully low errors (0.00 % of Mean error, 6.59 % of standard deviation of error, 10.83 % of 90 % of confidence interval, and 6.58 % of Mean RMS of 80 bundles data) when they are compared with correlations suggested by previous studies (CTD, UCTD, CTS, UCTS, and REH).