The Abelar pilot basin in Coruña (northwestern Spain) has been monitored for hydrological and hydrochemical data to assess the effects of eucalyptus plantation and manure applications on water resources, water quality, and nitrate contamination. Here, we report the machine learning analysis of hydrological and hydrochemical data from the Abelar basin. K-means cluster analysis (CA) is used to relate nitrate concentrations at the outlet of the basin with daily interflows and groundwater flows calculated with a hydrological balance. CA identifies three linearly separable clusters. Times series Gaussian process regression (TS-GPR) is employed to predict surface water nitrate concentration by incorporating hydrological variables as additional input parameters using a time series shifting. TS-GPR allows modelling nitrate concentrations based on shifted interflows and groundwater flows and chemical concentrations with R² = 0.82 and 0.80 for training and testing, respectively. Groundwater flow from five days prior to the current date, $Q_{g5}$, is the most important input parameter of the TS-GPR model. Interaction effects between the variables are found. TS-GPR validation with recent data provides results consistent with those of testing (R² = 0.85). Model inspection by permutation feature importance and partial dependence plots shows interactions between $Q_{g5}$ and Cl, and between Ca and Mg.

Groundwater flow and transport models are essential tools for assessing and quantifying the migration of organic contaminants at polluted sites. Uncertainties in the hydrodynamic and transport parameters of the aquifer have a significant effect on model predictions. Uncertainties can be quantified with advanced sensitivity methods such as Sobol’s High Dimensional Model Reduction (HDMR) and Variogram Analysis of Response Surfaces (VARS). Here we present the application of VARS and HDMR to assess the global sensitivities of the outputs of a transient groundwater flow model of the Gállego alluvial aquifer which is located downstream of the Sardas landfill in Huesca (Spain). The aquifer is subject to the tidal effects caused by the daily oscillations of the water level in the Sabiñánigo reservoir. Global sensitivities are analyzed for hydraulic heads, aquifer/reservoir fluxes, groundwater Darcy velocity, and hydraulic head calibration metrics. Input parameters include aquifer hydraulic conductivities and specific storage, aquitard vertical hydraulic conductivities, and boundary inflows and conductances. VARS, HDMR, and graphical methods agree to identify the most influential parameters, which for most of the outputs are the hydraulic conductivities of the zones closest to the landfill, the vertical hydraulic conductivity of the most permeable zones of the aquitard, and the boundary inflow coming from the landfill. The sensitivity of heads and aquifer/reservoir fluxes with respect to specific storage change with time. The aquifer/reservoir flux when the reservoir level is high shows interactions between specific storage and aquitard conductivity. VARS and HDMR parameter rankings are similar for the most influential parameters. However, there are discrepancies for the less relevant parameters. The efficiency of VARS was demonstrated by achieving stable results with a relatively small number of simulations.

The Abelar pilot basin in Coruña (Northwestern Spain) has been monitored for hydrological and hydrochemical data to assess the effects of eucalyptus plantation and manure applications on water resources, water quality and nitrate contamination. Here we report the machine learning analysis of hydrological and hydrochemical data from the Abelar basin. K-means Cluster Analysis (CA) is used to relate nitrate concentrations at the outlet of the basin with daily interflows and groundwater flows calculated with a hydrological balance. CA identifies three linearly separable clusters. Times Series Gaussian Process Regression (TS-GPR) is employed to predict surface water nitrate concentration by incorporating hydrological variables as additional input parameters using a time series shifting. TS-GPR allows modeling nitrate concentrations based on shifted interflows and groundwater flows and chemical concentrations with R2 = 0.82 and 0.80 for training and testing, respectively. Groundwater flow from five days prior to the current date, Qg5, is the most important input parameter of the TS-GPR model. Interaction effects between the variables are found. TS-GPR validation with recent data provides results consistent with those of testing (R2 = 0.85). Model inspection by permutation feature importance and partial dependence plots shows interactions between Qg5 and Cl, and between Ca and Mg.

Groundwater flow and transport models are essential tools for assessing and quantifying the mi-gration of organic contaminants at polluted sites. Uncertainties in the hydrodynamic and transport parameters of the aquifer have a significant effect on model predictions. Uncertainties can be quantified with advanced sensitivity methods such as Sobol’s High Dimensional Model Reduction (HDMR) and Variogram Analysis of Response Surfaces (VARS). Here we present the application of VARS and HDMR to assess the global sensitivities of the outputs of a transient groundwater flow model of the Gállego alluvial aquifer which is located downstream the Sardas landfill in Huesca (Spain). The aquifer is subject to the tidal effects caused by the daily oscillations of the water level in the Sabiñánigo reservoir. Global sensitivities are analysed for hydraulic heads, aquifer/reservoir fluxes, groundwater Darcy velocity and hydraulic head calibration metrics. Input parameters in-clude aquifer hydraulic conductivities and specific storage, aquitard vertical hydraulic conductiv-ities, and boundary inflows and conductances. VARS, HDMR and graphical methods agree to identify the most influential parameters which for most of the outputs are the hydraulic conduc-tivities of the zones closest to the landfill, the vertical hydraulic conductivity of the most permeable zones of the aquitard, and the boundary inflow coming from the landfill. The sensitivity of heads and aquifer/reservoir fluxes with respect to specific storage change with time. The aquifer/reservoir flux when the reservoir level is high shows interactions between specific storage and aquitard conductivity. VARS and HDMR parameter rankings are similar for the most influential parameters. However, there are discrepancies for the less relevant parameters. The efficiency of VARS was demonstrated by achieving stable results with a relatively small number of simulations.

The assessment of the long-term performance of the engineered barrier system (EBS) of a high-level radioactive waste (HLW) deep geological repository requires the use of high-fidelity reactive transport models. The EBS in a HLW repository includes: the canister, the compacted bentonite buffer and the concrete liner. Artificial intelligence and machine learning methods (ML) are growing at a very fast pace and have been used for: a) Accelerating numerical simulations, b) Addressing multiscale and multiphysics couplings, and c) Uncertainty quantification and sensitivity analyses. Here we present high-fidelity models and ML methods to simulate steel canister corrosion, corrosion products and their interactions with compacted bentonite. Metamodels and surrogate models provide approximate and efficient solutions which emulate the high-fidelity reactive transport simulations and can reduce significantly the CPU times. The high-fidelity model was calibrated with data from the FeMo corrosion test performed by CIEMAT/UAM under isothermal and saturated conditions for 15 years. The FeMo test consists of 6 stainless-steel sinters surrounded by Fe powder emplaced in holes drilled in a FEBEX bentonite block. The bentonite block was hydrated with granitic water through the sinters by using 6 syringes. Two different particle sizes (64 and 450 µm) were used in Fe powder of the FeMo tests. Model results show that pH increases to 9.5 and magnetite is the main corrosion product. Siderite, greenalite and saponite-Mg also precipitate at the Fe powder/bentonite interface. A metamodel has been developed for a geochemical system with interactions of steel/bentonite and precipitation of corrosion products. The system includes 3 primary dissolved species (Fe2+, H+ and O2aq), 2 aqueous complexes (OH- and H2aq) and magnetite. A set of 5000 data were sampled with a Latin Hyper Cube (LHC) sequence. Batch simulations were performed with CORE2Dv5 for 5000 data with the following 3 inputs: Fe, H and O2. Outputs include aqueous primary concentrations, aqueous secondary concentrations, magnetite, pH and Eh. The metamodel is based on Gaussian Processes and Random Forests for defining two groups corresponding to pH > 9 and pH ≤ 9. The metamodel provides excellent results for most of the output variables. Working with log for concentrations of H+, OH- and O2 improves significantly the results for H and O2. When the metamodel is trained by working with concentrations of dissolved Fe, the validation results show some negative concentrations. On the other hand, when the metamodel is trained by working with the logarithm of the concentrations of dissolved Fe, the predicted validation concentrations are always positive, but the metrics of the validation are slightly worse. The accuracy of the metamodel is significantly improved for pH by defining two groups, one for pH ≤ 9 and another for pH > 9.Acknowledgements: This research was funded by ENRESA within Work Package ACED of EURAD (Grant Agreement nº 847593), within WP HERMES of EURAD-2 (Grant Agreement nº 101166718) and Project PID2023-153202OB-I00 funded by Spanish Ministry of Science and Innovation We acknowledge the contributions of CIEMAT and UAM who performed FeMo tests and provided the experimental data.