A development of artificial neural network for the rapid assessment of subsurface contamination

Abstract

Contaminant transport in groundwater is a major environmental concern in many countries. It is desirable for users and regulatory agencies to have a reliable and convenient screening method to estimate the contaminant migration. Such a tool must allow users to identify critical sites and prioritize the clean-up and further investigative efforts. This study has investigated the feasibility of using an artificial neural network (ANN) as a quick predictor. A feedforward backpropagation neural network (FBNN) was used to train a data set simulated from the Visual MODFLOW program. The data set contained spatial and temporal variations of concentrations of one contaminant and the specified groundwater model parameters including hydraulic conductivity, longitudinal dispersivity, constant head difference, and recharge rate at the source of contamination under several hypothetical conditions. 1- and 2-hidden-layer FBNNs with various numbers of neurons were trained. Levenberg-Marquardt optimization with Bayesian regularization was found to be the best training algorithm for this training data set because of its capability of easing overfitting problem and improving generalization of the network. Data preprocessing prior training the network is also a crucial step for improving accuracy of prediction. The network with 2 hidden layers showed more accurate prediction than that with 1 hidden layer as indicated by the lower mean square error of the validation data set. The result indicates that this tool can be used to predict the concentration of contaminant as a function of time for given locations and model parameters.