Artificial Neural Network Code for Prediction of Friction Factor

Abstract

An artificial neural network code is introduced for the prediction of a single-variable output in a fluid flow problem. Using opensource Python code, this neural network is optimally designed for the prediction of frictional loss through pipe bends when a ground-truth simulation data is available for the model. The model is used to predict Darcy friction factor for a 90-degree and a 180-degree bend. Incompressible flow data used here is modelled using the Reynolds averaged Navier Stoke’s equations coupled with Reynolds stress model for the turbulence variables. Subsequently the neural network is trained and tested. It is found to perform with a mean square error of 0.0023 and a mean averaged error of 0.0355, along with an association strength of 0.9964 between the input and output variables. This is for input parameters such as Reynolds number between 10 and 0.1 million and curvature ratio between 0.01 and 0.2. With this we hope that the method will aid in supplementing computational analyses with its superior learning capabilities and also reduce the computational cost for such problems.