Artificial Neural Network (ANN) For Evaluating Permeability Decline in Permeable Reactive Barrier (PRB)

Environ. Process. (2015) 2:291–307 DOI 10.1007/s40710-015-0076-4 O R I G I N A L A RT I C L E Artificial Neural Network (ANN) For Evaluating Permeability Decline in Permeable Reactive Barrier (PRB) Umarat Santisukkasaem 1 & Fehintola Olawuyi 1 & Peter Oye 1 & Diganta B. Das 1 Received: 20 November 2014 / Accepted: 24 March 2015 / Published online: 18 April 2015 # Springer International Publishing Switzerland 2015 Abstract Artificial neural networks (ANNs) were developed which enable evaluation of longterm permeability losses that occur in permeable reactive barriers (PRBs) used in groundwater remediation. The network architectures consist of non-changing input and output layer(s) while the optimal hidden layer types and structures were determined through trial-and-error. Fluid residence time within the PRB, pressure drop, inlet volumetric flow rate, dynamic viscosity of fluid, average porosity, average particle size and the length of the reactor were selected as the input parameters to estimate the output parameter, namely, permeability. Of all experimental data available for each ANN structure, 70 % was used for training, 15 % for validation and the remaining 15 % for testing the ANN. The ANN structures were developed using a combination of soft computing techniques and mathematical association of varying physical parameters. Predictions obtained from the optimized ANN structures were compared with linear and non-linear regression models to assess their performance. The results indicate that ANN performs significantly better than the regression models and ANN modelling is a promising tool for the simulation and assessment of the permeability decline in PRBs. Keywords Artificial neural network (ANN) . Permeability loss . Permeable reactive barrier (PRB) . Zero-valent iron (ZVI) . Mineral precipitation 1 Introduction Permeable reactive barrier (PRB) is a well-known technology for groundwater treatment (U.S. EPA 2002; Das 2002; Chandrappa and Das 2012, 2014). It is a passive in-situ treatment wall (porous) filled with a reactive material and installed perpendicular to the groundwater flow in the subsurface (Fig. 1). The PRBs capture the contaminant plumes, break down the * Diganta B. Das 1 Department of Chemical Engineering, Loughborough University, Loughborough LE11 3TU Leicestershire, UK 292 U. Santisukkasaem et al. Contaminant plume PRB Treated water Groundwater flow Fig. 1 Schematic diagram showing contaminated groundwater flowing through a PRB contaminants and release the treated water into the surroundings. The contaminants are chemically, physically and/or biologically treated through the main processes of immobilization and transformation depending on the type of the reactive material. Conventionally, zerovalent iron (ZVI) is the reactive material used in PRBs but other materials such as surfactantmodified zeolites and peat moss have also been used (Chandrappa and Das 2012, 2014). ZVI has been widely used in PRBs as it is a highly reactive material and is suitable for treatment of various kinds of contaminants, i.e., heavy metals and hydrocarbons (Scherer et al. 2000; Vignola et al. 2011). One of the main limitations of this technology has to do with somewhat unpredictable longevity of the treatment system. This is mainly due to the intricate physicochemical processes, which occur in the PRB during the treatment process, one of which is permeability losses (Phillips et al. 2003; Li et al. 2005; Ruhl et al. 2011; Wilkin et al. 2014). Permeability decline of ZVI in PRB primarily takes place due to mineral precipitation (Liang et al. 2003; Hosseini et al. 2011). Apart from participating in reactions that are capable of degrading the contaminants, ZVI will also react with dissolved oxygen and other mineral constituents in the groundwater (mainly carbonates) as well as the water itself to form iron hydroxides. The formation of these secondary precipitates produces a coating on the ZVI particles surface and clogs the pores in the ZVI barrier. As a result, a decline in the porosity, and therefore, permeability of the ZVI barrier occurs, and the access of the contaminants to ZVI becomes constrained. As PRB is a passive treatment system, there are no additional forces that drive the flow of contaminant plumes through the PRB (Das 2005; Chandrappa and Das 2014). Therefore, the contaminated groundwater may bypass if the barrier is blocked significantly due to the reduction in permeability, and thus, the main function of the PRB, i.e., contaminant treatment, may be lost. As such, it is important to have an idea of the permeability loss over time for particular design of a PRB. It is evident from several literatures (e.g., Wilkin et al. 2002; Reardon 2005; Johnson et al. 2008) that the permeability loss occurs after the PRB has been installed in the subsurface. However, the data collected from field scale are difficult to use directly to determine the long-term permeability losses due to the fact that there are many uncontrolled parameters that affect the process. On the other hand, due to the time limitations, the data collection which is prolonged for many years in laboratory scale studies is often not practical. In principle, an approach based on computational fluid dynamics (CFD) may be employed to determine permeability losses (e.g., see, Liu et al. 2011) and the mathematical modelling should be capable of determining if there is a decline in the permeability (Jeen et al. 2012). However, the CFD tools generally require complex solution procedures for the governing equations for mass transport and fluid flow besides any other constitutive equations ANNs for Evaluating Permeability Decline in PRB 293 for changes in the physico-chemical properties of the PRB. The CFD models generally assume that the in-situ processes in the PRB can be described by well-defined parameters and governing equations; however, this is not necessarily the case. In a different context, it can be seen that several researchers have applied the artificial neural networks (ANNs) in predicting the permeability for porous domains (e.g., oil reservoirs, membrane filters for water treatment) and have concluded that ANN performed well and provided accurate results (Afshari et al. 2014; Inthata et al. 2013; Karimpouli et al. 2010). The ANN is a computational tool composed of simple elements working in parallel commonly known as neurons (Hanspal et al. 2013), which imitate the workings of the human brain and nervous system. The neurons are grouped into three interconnected layers (the input, hidden and output layers) (Modrogan et al. 2010; Deka and Quddus 2014). It is used as an alternative to logistic regression, which is a statistical technique with the same goal of predicting an outcome variable based on pre-determined set of input parameters. However, ANN is not a physically based approach and it relies on the network’s ability to understand given information and/or outputs from physical (...truncated)