A multi-output descriptive neural network for estimation of scour geometry downstream from hydraulic structures

Several researchers have attempted to estimate the maximum depth and location of local scour, particularly, based on conventional regression analysis. Many of these equations in the literature failed to estimate the scour depths satisfactorily. This study presents explicit formulation extracted from a multi-output descriptive neural network (DNN), which estimates both the depth and location of maximum scour. The DNN method extracts rules (information) conveyed from input layer to output layer of a NN consisting two outputs. The present DNN results are compared to non-linear and linear regression equations derived by the author and selected other empirical equations available in the literature. The results show that the proposed DNN estimates the maximum-scour depth and its location in strict agreement with the measured ones (R2 = 0.819 and 0.907, respectively), and dominantly better than the other equations (R2 = 0.687 and 0.706 being the highest results for dm and for xm, respectively). This study shows that the explicit formulation extracted from DNN can replace the conventional regression equations with much more accuracy.