Solute transport and environmental impacts in seawater-flooded soils: Insights from experiments, numerical modeling, and machine learning

• Both freshwater and seawater flooding increased concentrations of P, TP, and NH4-N in porewater samples. • The increase in porewater concentrations of P, TP, NH 4 -N, K, and Ca was more pronounced in seawater-flooded soils. • Seawater flooding led to increased adsorption of Na, Mg, and K to soil particles. • Hydrus-1D performed well in simulating Na and Mg transport in saturated soils. • Machine learning models are efficient in simulating the transport of various solutes. Seawater flooding salinizes soil, contaminates freshwater resources, and disrupts the optimal balance of macronutrients in soils, limiting the productivity of coastal agricultural lands. This study investigated the solubility and transport of cations and nutrients from calcareous soils in response to varying durations of seawater flooding. A series of column experiments was conducted by flooding two of the most common agricultural soils in South Florida (Biscayne and Krome), using two water sources (freshwater and seawater) for four different flooding durations: 1, 7, 14, and 28 days. A total of 540 porewater samples were collected from three distinct depths of the flooded soil columns and were analyzed for selected cations (Na + , Mg 2+ , K + , Ca 2+ ) and nutrients (P, TP, NO 3 -N, and NH 4 -N). A numerical hydrologic model, Hydrus-1D, and three data-driven machine learning (ML) models—Decision Tree (DT), Random Forest (RF), and Extreme Gradient Boost (XGB) were developed using experimental data to simulate the transport of cations and nutrients under saturated conditions. Results from the flooding experiment revealed a substantial release of cations, such as Ca 2+ , from seawater-flooded Biscayne and Krome soils. Compared to freshwater flooded soils, concentrations of Na + and Mg 2+ increased by a factor of 98 and 30 in Biscayne soil, and by 268 and 81 in Krome soil, respectively. Seawater flooding also elevated porewater concentrations of P, TP, and NH 4 -N by 82%, 56%, and 86% in Biscayne and by 14%, 7%, and 746% in Krome soils, respectively. The Hydrus-1D model effectively simulated the concentrations of Na + and Mg 2+ in porewater samples from seawater flooded Krome soils with r 2 values of 0.77 and 0.85 and RMSE of 0.44 and 0.17mg cm -3, and with r 2 values of 0.93 and 0.75 and RMSE values of 0.09 and 0.17 mg cm -3 in porewater samples collected from seawater flooded Biscayne soils, respectively. The data-driven ML models (DT, RF, and XGB) outperformed Hydrus-1D in simulating the concentrations of all cations and nutrients under saturated conditions, except NO 3 -N. Among the ML models, the XGB performed best in simulating P, TP, NH 4 -N, Ca, K, Na, and Mg concentrations with r 2 and RMSE values ranging from 0.53 to 0.84 and 0.3 to 1145 mg L -1 , respectively. Overall, ML models could be considered viable options in simulating the concentration and dynamics of cations and nutrients, but extensive data requirements for training and testing these models could be a limitation.