Deep Learning for Groundwater Prediction

This thesis explores the integration of advanced machine learning techniques, particularly deep learning, in enhancing groundwater prediction models. The primary focus is on developing new surrogate models that leverage deep neural networks for simulating groundwater flow, bridging the gap between traditional hydrological methods and contemporary data science approaches. The research journey begins with the application of synthetic data and computer vision techniques and progressively advances towards handling sparse data and real-world scenarios. The thesis comprises four key papers, each contributing to the development of machine learning models for groundwater prediction. These models include convolutional encoder-decoder networks (Attention U-Net and U-Net integrated with Vision Transformer) for accurate steady-state response prediction, the DeepONet framework for generalized groundwater flow modeling under data-sparse scenarios, and finally spatial-temporal graph neural networks for long-term forecasting of groundwater levels. The research demonstrates the ability of these models to handle complex hydrological systems, predict accurately under varying conditions, and efficiently process both high-dimensional inputs and sparse data.

Overall, this thesis contributes to the field of hydrology by establishing advanced machine learning models as viable alternatives for predictive groundwater level modeling, particularly noted for their accuracy, computational efficiency, and adaptability to diverse scenarios. The findings pave the way for future research, focusing on applying these models to larger and more complex datasets for practical use in groundwater management and decision-making.