Assessment of parameters of groundwater models using artificial neural networks

Abstract

The paper investigates the possibility of solving the inverse problem using artificial neural networks. To demonstrate the approach, an example is considered consisting of an analytical model of the transfer of pollutants from a point source to a stationary flow field. The model was used to model the behavior of groundwater systems at different values of the dispersion coefficient. Next, a set of controlled multilayer neural networks of direct propagation was trained to evaluate, determine, and select a parameter corresponding to given concentration histories. The results obtained in the work showed satisfactory accuracy of neural network estimates, which confirms the stability of the approach to data analysis in field experiments. When training four artificial neural networks of a controlled, multilayer and direct type, it was found that each of them specialized in a wide range of values. This led to more accurate predictions compared to the case of training a single network over the entire range of values. In addition, the paper shows the ability of a neural network to identify the dispersion parameter at a given concentration under the influence of “noise”. An analysis of the topologies of the applied neural networks has established that the presence of 10 hidden nodes is sufficient to ensure a satisfactory level of calculation accuracy.