Machine learning and parametrisation of multi-cell structures of secondary circulation in a tight open channel bend using LES

Abstract

Large eddy simulations of an open channel bend are performed at a variety of water depths and flow rates. The results at several cross sections are decomposed into sub-cells of secondary circulation using clusters of instantaneous vortices. The strength and position of the sub-cells are then modelled using decision trees, multiple linear regression, multi-layer perceptrons, and adaptive neuro-fuzzy inference systems to obtain parametric models of secondary circulation development in a channel bend. The development of individual cells and total circulation is shown for an arbitrary flow condition using the model, as well as the dependence of all the circulation output variables on the input parameters of aspect ratio and Froude number. The positions of the sub-cells (but not their circulations) are largely independent of the Froude number, and the cross-stream position of the centre cell is found to behave linearly. The model with the best performance across all predicted variables is the ANFIS model without classification.