Machine learning assisted, laboratory and numerical study of saltwater intrusion in stratified and fractured coastal aquifers

Abstract

Coastal aquifers are the main source of freshwater for billions of people around the globe, thus identifying the main mechanisms in which aquifer salinization occurs constitutes an important environmental challenge. The main objective of this thesis was to expand the current knowledge on the impact of coastal aquifer heterogeneity on the saltwater intrusion (SWI) hydrodynamics. A series of laboratory experiments, supplemented by numerical simulations, were employed to study the fundamental mechanisms of SWI in stratified and fractured hydrological systems. The experimental setup consisted of a thin sandbox where the aquifers’ porous medium was recreated using clear glass beads of varying diameters, while coloured saltwater was utilized to vividly visualize the interface between the two liquids. The numerical code SUTRA was used for the modelling of groundwater flow. Three fundamental variables were utilized to quantify the extent of saline intrusion in the investigated aquifers, the toe length (TL) of the intruding wedge, the width of the mixing zone (WMZ) and the angle of intrusion (AOI).Obtaining saline concentration fields through the digital processing of sandbox images has become an established practice in laboratory investigations of SWI. To better facilitate the goals of this study, two distinct approaches of optimizing this experimental process were developed. The first technique employed a novel image filtering procedure alongside the predictive power of multiple machine learning algorithms. Its application diminished the necessity of conducting experimental calibrations for each investigated aquifer, speeding-up the data acquisition procedure by up to 50 %. The methodology was applicable in both homogeneous and heterogeneous setups alike, while it was less prone to a series of previously documented experimental errors, outperforming the state-of-the-art image processing techniques currently applied in laboratory studies of SWI. The second approach investigated the impact of image resolution and image colour depth on the application of the traditional pixel wise regression methodology. The technique identified an optimum combination between the two variables that can potentially minimize the image processing time without compromising the quality of the recreated saltwater concentrations. The machine learning assisted approach was the one employed in the processing of all the acquired experimental SWI images presented in the following chapters of this thesis. Sandbox experiments and numerical simulations in stratified laboratory aquifers, revealed that the TL of the intruding wedge was affected by a series of heterogeneity variables such as the number of the various strata, and the permeability contrast between them, as well as the total aquifer transmissivity. Regarding the WMZ, whenever two layers with distinct permeabilities were located next to each other, the mixing zone widened in the less permeable layer while it became thinner for the more permeable one. The degree of this alteration was positively correlated to the permeability contrast of the adjacent strata. Moreover, changes in the applied boundary conditions led to the transient widening of the mixing zone, especially during saltwater retreat. Similar to what was established for the WMZ, milder angles of intrusion were observed in aquifers where more permeable layers overlaid lower permeability strata, and steeper AOIs in the inverse aquifer setups. Finally, the existence of a less permeable upper zone, led to the confinement of saltwater into the lowermost part of the aquifer; until a critical hydraulic head difference was applied to the system.The study of head-induced SWI in fractured aquifers indicated that high-permeability discontinuities can significantly affect saltwater hydrodynamics. In most experiments, the TL was positively correlated to the horizontal fracture’s distance from the landward aquifer boundary. Further numerical investigation revealed that this relationship is not as straightforward, instead the presence of a single discontinuity can either limit or exacerbate SWI, depending on the hydraulic gradient applied on the aquifer. Horizontal fractures determined the dimensions of the saline wedges not only in the horizontal but also in the vertical direction, while their impact in saline hydrodynamics was proportional to their length. Even though saline TL was not affected to the same extent by vertically oriented fractures, their presence led to substantial widening of the mixing zone. Lastly, in laboratory setups where the discontinuities were in direct proximity to the aquifer’s side boundaries; a distinct freshwater-saltwater interface, that deviated from the rest of the experimental cases, was identified.In the final part of the thesis the saltwater upconing mechanism in fractured coastal aquifers was investigated. A peristaltic pump was employed to abstract freshwater from sandbox aquifers containing horizontal discontinuities, until critical pumping, causing well salinization, was achieved. Critical abstraction rates were positively correlated to the fractures’ length, permeability, and distance from the sea, as well as the permeability of the surrounding porous medium. On the other hand, these pumping rates were negatively correlated to the hydraulic head difference applied in the system and the well’s depth. Under critical abstraction, fractured aquifers with similar permeability contrast between their discontinuities and the porous medium, but with different total transmissivities, demonstrated almost identical freshwater-saltwater interfaces. Finally, during the post-pumping saltwater retreat phase of the experiments, the presence of the discontinuities caused the segmentation of the saline front into distinct plumes that retreated separately towards the sea.

Date of Award	Dec 2022
Original language	English
Awarding Institution	Queen's University Belfast
Supervisor	Gerard Hamill (Supervisor) & Raymond Flynn (Supervisor)