Abstract
The escalating frequency and intensity of floods, driven by climate change, necessitate the development of rapid and accurate flood forecasting and hazard mapping systems. While traditional hydrodynamic models provide high-fidelity predictions, their immense computational cost is prohibitive for real-time applications and large-scale analyses. In response, data-driven surrogate models, particularly Graph Neural Networks (GNNs), have emerged as a promising, computationally efficient alternative. However, their widespread adoption is hindered by several critical challenges: the high cost of generating sufficient training data, a tendency to produce physically implausible predictions, a lack of model interpretability, and poor generalizability to new geographical domains.
The overarching objective of this dissertation is to systematically address these challenges through the development and evaluation of a suite of advanced deep learning frameworks. This is achieved by moving beyond conventional architectures to create surrogate models that are not only fast and accurate but also data-efficient, physically consistent, interpretable, and adaptable.
This dissertation makes three primary scientific contributions. First, to address the data-generation bottleneck, a Multi-Fidelity Graph Neural Network is proposed. This framework leverages a hierarchical learning strategy to effectively combine abundant, low-fidelity (coarse-resolution) simulation data with a sparse set of high-fidelity samples, dramatically reducing the computational cost of training without sacrificing predictive accuracy.
Second, to tackle the issues of physical inconsistency and a lack of transparency, HydroGraphNet, a novel physics-informed GNN, is introduced. This framework explicitly embeds mass conservation laws into the training objective to ensure physically realistic predictions. Furthermore, it pioneers the integration of Kolmogorov-Arnold Networks into a GNN for flood forecasting, a unique architectural innovation that greatly enhances model interpretability by revealing the learned functional relationships between environmental variables and flood dynamics.
Third, to overcome limitations in handling complex geometries and to enhance model transferability, this dissertation presents FloodForecaster, a framework built upon a novel Geometry-Informed Neural Operator (GINO). GINO synergistically combines the strengths of Graph and Fourier Neural Operators to adeptly model intricate local terrain features while efficiently capturing long-range global dependencies. Crucially, this framework incorporates a domain adaptation strategy that enables the model to generalize to new, data-scarce regions, thereby overcoming the critical challenge of domain shift.
Extensive validation of these frameworks on real-world fluvial and pluvial flood scenarios demonstrates their effectiveness. The proposed models are shown to be not only orders of magnitude faster than traditional numerical simulators but also more accurate, physically robust, and transferable than standard data-driven baselines. Collectively, the contributions of this dissertation represent a significant step toward the development of the reliable, scalable, and trustworthy deep learning-based tools required for next-generation flood warning and risk management systems.