Forecasting solute breakthrough curves through the unsaturated zone using artificial neural networks

Abstract

Effective groundwater management requires precise forecasting of the amount of contaminants intruding into groundwater from the surface. In this study, solute breakthrough curves throughout the unsaturated zone were predicted using artificial neural networks (ANNs), through numerical tests and through laboratory experiments. In the numerical tests, the applicability of the ANN model to the prediction of breakthrough curves was evaluated using synthetic data generated by a groundwater flow and transport model in a variably saturated media, HYDRUS-2D. The use of two ANNs, one for solute arrival times and the other for solute mass breakthroughs after the solute arrival time, was suggested in order to reduce the prediction error. The results showed that the network building process was essential in ANN model applications. The best ANN model gave a correlation coefficient value between target and output values of over 0.98. The sensitivity analysis of data forms for the network training demonstrated that regular breakthrough curves that contain a peak value can train the ANN model effectively. Then, the ANN model was verified using laboratory data obtained by tracer infiltration tests in a sand column. The overall results demonstrate that the ANN model can be an effective method for forecasting solute breakthrough curves through the unsaturated zone when hydraulic data are available.