Flow data forecasting for the junction flow using artificial neural network

The present study aims to predict the flow characteristics downstream of a cylinder, which is the result of junction flow using an Artificial Neural Network (ANN) algorithm. The training and test datasets were obtained through Particle Image Velocimetry (PIV) experiments. The experiments were conducted at Reynolds numbers Re = 1.5 x 10(3) and 4 x 10(3) based on the cylinder diameter (D) at dimensionless measurement heights (Z = h/D) of Z(1) = 0.06, Z(2) = 0.4, Z(3) = 0.8, and Z(4) = 1.6 respectively. While the X- and Y-coordinate and dimensionless measurement location (Z) variables are employed as inputs to the ANN model, the output variables are vorticity <omega >, streamwise velocity < u >, and transverse velocity < v >, which are derived from the time-averaged flow data. Modeling flow characteristics with easily obtainable independent variables without flow and physical properties was considered. Three various training algorithms such as Levenberg Marquardt (LM), Resilient Backpropagation (RP), and Scaled Conjugate Gradient (SCG) were employed to assess and compare their prediction performance. The results indicate that the LM learning algorithm outperforms the RP and SCG algorithms, especially at low Reynolds (Re) numbers. The ANN model, trained with the LM algorithm, exhibits significant success, achieving R = 0.9816 correlation coefficient (R), MAE = 2.4250 m/s Mean Absolute Error (MAE), and RMSE = 3.3541 m/s Root Mean Square Error (RMSE) for streamwise velocity < u > data. Notably, the LM algorithm for the testing process demonstrates the best predictions at Re = 1.5x10(3), yielding R = 0.9779, MAE = 2.7417 m/s, and RMSE = 3.7493 m/s. The ANN-LM model's patterns closely align with experimental results, affirming its accuracy, which proves that the prediction of time-averaged velocity data solely based on spatial coordinates as input can be achieved successfully.