Parameterization and prediction of nanoparticle transport in porous media: A reanalysis using artificial neural network

Abstract

The continuing rapid expansion of industrial and consumer processes based on nanoparticles (NP) necessitates a robust model for delineating their fate and transport in groundwater. An ability to reliably specify the full parameter set for prediction of NP transport using continuum models is crucial. In this paper we report the reanalysis of a data set of 493 published column experiment outcomes together with their continuum modeling results. Experimental properties were parameterized into 20 factors which are commonly available. They were then used to predict five key continuum model parameters as well as the effluent concentration via artificial neural network (ANN)-based correlations. The Partial Derivatives (PaD) technique and Monte Carlo method were used for the analysis of sensitivities and model-produced uncertainties, respectively. The outcomes shed light on several controversial relationships between the parameters, e.g., it was revealed that the trend of K-att with average pore water velocity was positive. The resulting correlations, despite being developed based on a "black-box" technique (ANN), were able to explain the effects of theoretical parameters such as critical deposition concentration (CDC), even though these parameters were not explicitly considered in the model. Porous media heterogeneity was considered as a parameter for the first time and showed sensitivities higher than those of dispersivity. The model performance was validated well against subsets of the experimental data and was compared with current models. The robustness of the correlation matrices was not completely satisfactory, since they failed to predict the experimental breakthrough curves (BTCs) at extreme values of ionic strengths. Plain Language Summary Models based on advection-dispersion-equation (ADE), have succeeded in describing a variety of nanoparticle (NP) transport mechanisms within subsurface porous media. These models are usually fitted against known observation data to obtain the unknown parameters. Nevertheless, the parameters determined in this way cannot be used for a new problem of the same type and again there exists the need for the data to calibrate the model parameters for a new problem. Black box models, such as artificial neural network (ANN), have been mostly, if not all the times, used in the same way of ADE models, i.e., single problem solver. In this paper we use the ability of ANN to develop a series of simple correlation matrices that can be easily used for the prediction of ADE parameters in the new problem without the need for additional calibration. Although comparisons between ANN model predictions and experimental data show that there is still further work to be done, our approach out-performs other comparable models and offers new insight into the complex interactions among the factors determining NP transport and fate in the environment.