3D HYDROSTRATIGRAPHIC MODELING USING SUPERVISED MACHINE LEARNING: CASE STUDY OF THE HIGH PLAINS AQUIFER, SW NEBRASKA

Heterogeneous aquifer-bearing zones can impact variable groundwater flow over a short distances. Such processes are difficult to simulate with groundwater modeling technique unless the model is supported by high-resolution 3D geology and hydrogeology data. This research uses a supervised machine learning approach to predict lithology for a hydrostratigraphic model that can be used to solve groundwater sustainability problems in the High Plains Aquifer of southwestern Nebraska. The model is trained using location and resistivity parameters to predict hydrostratigraphic characteristics. More than 2000 boreholes and 2717 km of Airborne Electromagnetic (AEM) survey data are used. The borehole data contains ~80 unique lithological terms that we grouped into 5 hydrostratigraphic units based on grain size, dominant lithology, assumed response to geophysical parameters, and hydrogeologic characteristics. The borehole and resistivity data are resampled into 200X200X1m using a gridding technique by Geoscene 3D software. Then the data is split into 70% training and 30% for validation to train and test a random forest model. Confusion matrices, precision, recall, and F1 score accuracy metrics are used for model performance evaluation. Random forest showed excellent performance with training F1 score of 97% and 91% testing accuracy in learning a co-linear relationship between location and resistivity vs. lithology. After hyperparameter tuning the training accuracy improved to 100% and testing accuracy of 95%. This method produces a high-resolution lithostratigraphic and hydrostratigraphic model at the resolution of the AEM survey. It fills gaps between widely spaced boreholes, producing a high-resolution 3D model that can be applied in any setting where AEM data is available.