Evaluating the feasibility of using artificial neural networks to predict hydrogeologic units in complex glacial deposits

The feasibility of using Multilayer Perceptron (MLP) to predict hydrogeologic units (HGUs) in complex glacial deposits is evaluated. The study area includes the Fraser-Whatcom Basin. Material descriptions from boreholes logs are standardized into HGUs using natural language processing techniques to reduce subjectively and improve automation. Three data selection alternatives are considered to evaluate the training and prediction capabilities of MLP. Block-model representations of the subsurface are created and the best geologic realization is verified against predictions using the K-nearest neighbours algorithm and geologic cross-sections from independent studies. Validation results show MLP predictions are typically more generalized but produce similar subsurface trends and can recreate confining units contributing to local artesian conditions. MLP appears to be a promising algorithm to solve multi-class classification for geologic modelling purposes. The workflow developed has the added benefit of being stochastic with the potential to generate multiple geologic realizations to account for uncertainty in geologic structure.