Paper Title
Predicting Exit Gradient of Water Retaining Structures: A Comparative Study of Five Machine Learning Techniques
Abstract
Numerical seepage analysis under hydraulic water retaining structures(HWRS) is utilized to accurately simulate
the seeping water impacts and determine seepage quantities. The most important seepage characteristic is the exit gradient.
Recently, the numerical simulation models are being directly or indirectly linked to other models or systems to implement a
design methodology incorporating optimization or reliability models. Since it may be computationally expensive to directly
link the numerical model to the other models, an efficient surrogate model based on machine learning technique could be
trained to accurately predict the numerical responses. The surrogate models could expeditiously imitate the numerical
models responses, even for scenarios and input data not utilized in the training process. This study concentrates on
comparing the performance of five different machine learning techniques used to develop surrogate models to predict the
exit gradient value for designing HWRS. The exit gradient data results from the numerical seepage modelling using SEEPW
/ Geo-Studio code. The machine learning techniques utilized in this study encompass the artificial neural network (ANN),
support vector machine (SVM), Multi-gens Genetic programming (MGP), Gaussian process regression (GPR) and Adaptive
Neuro-fuzzy inference system (ANFIS). The optimum parameters of each technique were systematically obtained based on
the Taguchi design of experiment (DOE) technique. The solution results demonstrated that the MGP could provide the best
exit gradient prediction for a certain parameter combinations. Also, based on the limited evaluations the worse machine
learning technique was identified as the SVM.
Keyword - Exit gradient; Machine learning technique; Taguchi DOE, Numerical seepage analysis