Date of Defense
12-11-2025 3:30 PM
Location
H1 - 0017B
Document Type
Dissertation Defense
Degree Name
Doctor of Philosophy in Civil Engineering
College
College of Engineering
Department
Civil and Environmental Engineering
First Advisor
Prof. Mohsen Sherif
Keywords
Seawater intrusion; Hydraulic barriers; Physical barriers, FEFLOW, Optimization, Machine learning, SHAP, Hyper-arid aquifers, Fujairah, UAE.
Abstract
Seawater intrusion (SWI) threatens the reliability of coastal groundwater especially in hyper-arid settings, where climatic stress and pumping accelerate salinization. This dissertation advances two complementary approaches to managing SWI: Part A optimizes mitigation measures, hydraulic (pumping/injection) and physical barriers (cutoff walls, subsurface dams) on benchmark models; Part B predicts SWI in the hyper-arid Fujairah (UAE) coastal aquifer using total dissolved solids (TDS) as a proxy, through machine learning-based models, spatially and dynamically. A bibliometric synthesis first maps the evolution of SWI models and mitigation strategies, identifying gaps that motivate the subsequent methodological developments. Part A employs the classical Henry problem to develop and test mitigation designs independent of site specifics. A FEFLOW–Python optimization workflow derives explicit design correlations for hydraulic barriers, including “negative” barriers that extract brackish water from the dispersion zone. Optimal extraction is near the center-bottom of the wedge; optimal injection depends on rate, with maximum retardation (~61%) when placed at the aquifer bottom near the coast, shifting toward the toe at lower rates. A second benchmark study couples a Python–FEFLOW simulator with machine-learning sensitivity analysis (Random Forest + SHAP) to evaluate physical barriers (cutoff walls, subsurface dams). Under uniform properties, cutoff walls reduce total dissolved salts by up to ~98% and seawater influx by ~76.5%, while subsurface dams achieve ~92% and ~81%, respectively; proximity to the saline boundary can induce stagnation zones and contour distortion, underscoring siting trade-offs. Part B targets the Fujairah coastal aquifer (UAE). Using six hydrogeologic predictors, fifteen algorithms are benchmarked for TDS-based SWI assessment: LightGBM attains R² ≈ 0.957 for prediction, and Gradient Boosting/CatBoost reach ~97.6% accuracy (AUC ≈ 0.999) for classification. Derived empirical equations provide practical screening tools, with hydraulic head and distance from the coast emerging as dominant drivers. Subsequently, a time-dependent deep-learning program forecasts daily TDS from >14,000 records at 16 wells, comparing FFNN, LSTM, Transformer, and a hybrid LSTM-FFNN. After KNN imputation and normalization, an attention-augmented LSTM provides the most reliable temporal forecasts and sustains the highest accuracy across seasonal regimes (including high-stress summer months) and spatial zones (coastal–inland gradients) (MAE ≈ 401 mg/L; R² ≈ 0.983). Building on Parts A and B, future research should (i) translate the benchmark-derived design rules to site-specific, 3D variable-density models that include realistic heterogeneity, anisotropy, tidal forcing, sea-level rise, and temperature/viscosity effects, and validate them via controlled field pilots (instrumented pumping/injection tests, multi-depth EC/TDS logs, and time-lapse geophysics); (ii) couple process models with data assimilation for real-time state estimation and barrier control; (iii) advance prediction by fusing additional drivers (pumping schedules, tides, land use); (iv) develop active-learning strategies to optimize monitoring network design and target new wells/sensors where forecasts are most uncertain; and (v) generalize and transfer the Fujairah-trained pipelines to other hyper-arid coasts via domain adaptation and few-shot fine-tuning, culminating in a “digital twin” for adaptive SWI management.
Included in
INTEGRATED OPTIMIZATION AND DATA-DRIVEN MODELING FOR SEAWATER INTRUSION MITIGATION AND PREDICTION
H1 - 0017B
Seawater intrusion (SWI) threatens the reliability of coastal groundwater especially in hyper-arid settings, where climatic stress and pumping accelerate salinization. This dissertation advances two complementary approaches to managing SWI: Part A optimizes mitigation measures, hydraulic (pumping/injection) and physical barriers (cutoff walls, subsurface dams) on benchmark models; Part B predicts SWI in the hyper-arid Fujairah (UAE) coastal aquifer using total dissolved solids (TDS) as a proxy, through machine learning-based models, spatially and dynamically. A bibliometric synthesis first maps the evolution of SWI models and mitigation strategies, identifying gaps that motivate the subsequent methodological developments. Part A employs the classical Henry problem to develop and test mitigation designs independent of site specifics. A FEFLOW–Python optimization workflow derives explicit design correlations for hydraulic barriers, including “negative” barriers that extract brackish water from the dispersion zone. Optimal extraction is near the center-bottom of the wedge; optimal injection depends on rate, with maximum retardation (~61%) when placed at the aquifer bottom near the coast, shifting toward the toe at lower rates. A second benchmark study couples a Python–FEFLOW simulator with machine-learning sensitivity analysis (Random Forest + SHAP) to evaluate physical barriers (cutoff walls, subsurface dams). Under uniform properties, cutoff walls reduce total dissolved salts by up to ~98% and seawater influx by ~76.5%, while subsurface dams achieve ~92% and ~81%, respectively; proximity to the saline boundary can induce stagnation zones and contour distortion, underscoring siting trade-offs. Part B targets the Fujairah coastal aquifer (UAE). Using six hydrogeologic predictors, fifteen algorithms are benchmarked for TDS-based SWI assessment: LightGBM attains R² ≈ 0.957 for prediction, and Gradient Boosting/CatBoost reach ~97.6% accuracy (AUC ≈ 0.999) for classification. Derived empirical equations provide practical screening tools, with hydraulic head and distance from the coast emerging as dominant drivers. Subsequently, a time-dependent deep-learning program forecasts daily TDS from >14,000 records at 16 wells, comparing FFNN, LSTM, Transformer, and a hybrid LSTM-FFNN. After KNN imputation and normalization, an attention-augmented LSTM provides the most reliable temporal forecasts and sustains the highest accuracy across seasonal regimes (including high-stress summer months) and spatial zones (coastal–inland gradients) (MAE ≈ 401 mg/L; R² ≈ 0.983). Building on Parts A and B, future research should (i) translate the benchmark-derived design rules to site-specific, 3D variable-density models that include realistic heterogeneity, anisotropy, tidal forcing, sea-level rise, and temperature/viscosity effects, and validate them via controlled field pilots (instrumented pumping/injection tests, multi-depth EC/TDS logs, and time-lapse geophysics); (ii) couple process models with data assimilation for real-time state estimation and barrier control; (iii) advance prediction by fusing additional drivers (pumping schedules, tides, land use); (iv) develop active-learning strategies to optimize monitoring network design and target new wells/sensors where forecasts are most uncertain; and (v) generalize and transfer the Fujairah-trained pipelines to other hyper-arid coasts via domain adaptation and few-shot fine-tuning, culminating in a “digital twin” for adaptive SWI management.