Key points are not available for this paper at this time.
This study outlines a systematic workflow for developing a 3D geological model with the goal of converting it into a groundwater flow model to simulate groundwater dynamics in the lower Spree catchment in Berlin/Brandenburg, Germany. The flow model is constructed by employing the finite-difference method to discretize groundwater flow equations, focusing on investigating saltwater up-coning resulting from changes in recharge or increased pumping. For this aim geological cross-section profiles and borehole data are used to construct a 3D geological model covering 1100 km with depths up to 350 m. This model serves as a solid base for the subsequent 3D flow model with fine spatial resolution (100 m horizontally and 5 m vertically) and monthly temporal resolution. The foundation of the geological model hinges on the stratigraphic order and addresses structural complexities such as faults, folds, dip angles, azimuth, and extensions. Challenges arise from outdated data, requiring meticulous preprocessing and cleaning, especially for larger areas with intricate structures and lenses. Specific challenges involve multiple boreholes sharing identical coordinates with varying lithological and stratigraphical descriptions for the same depth, as well as the repetition of layers. Addressing these issues requires careful preprocessing and cleaning. The study explores interpolation techniques, including Inverse Distance Weighting and Natural Neighbor algorithms, commonly used in geospatial analysis. Challenges related to these methods are discussed, emphasizing difficulties in dealing with shallow boreholes, missing stratigraphy, and ensuring layer continuity, affecting the overall reliability of interpolation results. To overcome these challenges, a Multi-Layer Perceptron (MLP) machine learning classifier is introduced. This classifier learns the hierarchical order of lithological information, seamlessly integrating it into the geological model. The MLP classifier is trained on preprocessed data, utilizes a dataset split into 85% for model training and 15% for validation, achieving a validation score of 73%. The geological model is then converted into a numerical mesh for the groundwater flow model. Hydraulic parameters, including hydraulic conductivity, porosity, specific storage, and specific yield, are estimated using empirical formulas and correlation sheets such as the Hazen and Bayer methods, which involve the determination of effective grain size (D10 and D30). Statistical algorithms aid in identifying and assigning hydraulic parameters related to the dominant material group to each flow cell. In instances where two material groups exhibit the same dominance, average values of hydraulic parameters are calculated. In conclusion, the study highlights the utilization of advanced techniques, including machine learning, alongside statistical methods, becomes imperative to solve complex geological settings, ensuring a more accurate and reliable representation of subsurface properties for groundwater flow models.
Abdelrahman et al. (Fri,) studied this question.