Assessment of artificial neural network models based on the simulation of groundwater contaminant transport

Abstract

The characterization of contaminant transport in subsurface environments is a pre-requisite for sustainable groundwater use and management. Various analytical and numerical models are generally utilized for such characterization. Analytical models are used for solving simple and idealized pollution transport problems, while numerical models deal with real-world pollution transport simulations. However, in situations where any of the aquifer parameters (such as geological properties, boundary conditions, initial conditions, etc.) are not explicitly defined, use of these models becomes redundant. Simulation of pollution transport under such scenarios becomes challenging. To tackle such problems, researchers generally make use of artificial neural network (ANN) models. Existing literature review reveals that feed-forward backpropagation (FFBNN) models are the most commonly used training method while the applicability of other ANN models has not been appropriately explored. In this study, various ANN models, encompassing supervised and unsupervised neural networks, like, cascade-forward backpropagation (CFBNN), FFBNN, radial basis function (RBFNN), exact radial basis function (ERBFNN) and generalized regression (GRNN), have been developed, and their performances are compared for transport simulation of a conservative pollutant in a two-dimensional hypothetical aquifer. The models reported have pollution injection rates and locations of the sources as input data and the pollution concentrations measured in some water supply wells (or monitoring wells) are considered as target data. Stability of the developed models has been validated by sensitivity analyses. This study reveals that alternative ANN models (other than the ones already reported in literature) can be reliable for simulation of pollutant transport in groundwater systems.