Hydrodynamic flood models play a central role in understanding the spatiotemporal evolution of floods for effective disaster management. Current implementations are limited, as physics-based numerical models are constrained by slow inference times, and data-driven models lack interpretability and physical consistency. Graph Neural Networks (GNNs) provide a physically meaningful framework to represent unstructured spatial domains as graphs, mirroring the node-link structure of numerical models. Unlike other deep learning methods, GNNs require no complex spatial data transformation and can seamlessly integrate physics-based priors. This review synthesizes forty nine peer-reviewed studies on GNN applications in flood modeling, identifying key gaps in physics incorporation, scalability, and cross-domain generalization. Drawing on advances from related fields where GNNs have excelled, we outline strategies to overcome these limitations. GNNs hold significant potential as next-generation flood models by combining physical realism, computational efficiency, and adaptability for real-time flood prediction and risk management.

Advances in machine learning have improved rainfall–runoff prediction, but integrating these capabilities with interpretable, process-based model structures remains an open challenge. This study seeks a middle ground between conceptual hydrological modelling and modern data-driven prediction by automating model structure discovery while retaining interpretability and hydrological consistency. We introduce HydroXGP (Hydrological Model eXplorer using Genetic Programming) to reduce reliance on legacy model choices and subjective a priori structural assumptions by learning model structure directly from data during optimization. The framework evolves models from a modular, process-based component library adapted from the open-source MARRMoT toolbox, making the search process-informed through reusable flux and routing components and hydrologically constrained by enforcing mass balance during simulation. GP jointly optimizes model structure and parameters, while each discovered model remains explicit and interpretable. Evaluation on the CAMELS-US dataset showed that a proof-of-concept test in a snow-dominated basin achieved strong performance over the out-of-sample test period (NSE=0.849), and, when transferred to a hydrologically similar basin, the learned structure retained useful predictive skill (NSE=0.724). Across 18 basins (one representative basin per U.S. hydrologic region), median NSE reached 0.702, generally outperforming the mean performance of an ensemble of conceptual benchmark models while remaining competitive with some regional LSTM baselines. The findings support a modelling perspective in which structural hypotheses can be revisited and adapted as new data, diagnostics, or environmental signals become available, enabling basin-specific configurations that evolve in response to changing information rather than relying on a single fixed template.

Physics-based hydrodynamic models are essential for accurate flood prediction but are computationally expensive, limiting their applicability for real-time forecasting and probabilistic analyses. Conversely, pure machine learning (ML) models offer both computational efficiency and accuracy but often lack interpretability. To address this gap, we propose SGUnet, a physics-informed ML model and a hybrid theory-guided data science approach, for rapid, high-resolution flood mapping. It utilizes a neural network with U-Net architecture and integrates subgrid-based coarse-grid hydrodynamic model predictions as initial estimates, upskilling them to achieve fine-grid model accuracy. Unlike traditional hydrodynamic models, the subgrid method embeds fine-scale topographic details within coarse-grid cells, enhancing both computational efficiency and predictive accuracy. SGUnet processes flood depth raster patches (512 × 512 pixels) and corresponding digital elevation models as inputs. It functions as a deep learning-based corrector, refining flood predictions from numerical simulators. Trained through supervised learning, SGUnet learns to correct deviations in coarse-grid predictions using fine-grid model outputs as target values. The model is evaluated across three large Australian watersheds—Wollombi, Chowilla, and Burnett River—using HEC-RAS flood simulations with subgrid formulation. SGUnet reduces root mean squared error by a factor of 4.5–5.3 compared to coarse-grid models, achieves a critical success index exceeding 0.9 for flood extent mapping, and delivers a 50x speed-up over fine-grid hydrodynamic models. Furthermore, SGUnet outperforms a state-of-the-art ML-based upskilling model in depth and extent predictions. By effectively correcting flood artifacts from coarse-grid models, SGUnet achieves near fine-grid accuracy with significantly reduced computational cost, demonstrating its potential for real-time flood risk assessment.

Environmental sustainability is a major concern for urban and rural development. Actors and stakeholders need economic, effective and efficient simulations in order to predict and evaluate the impact of development on the environment and the constraints that the environment imposes on development. Numerical simulation models are usually computation expensive and require expert knowledge. We consider the problem of hydrological modelling and simulation. With a training set consisting of pairs of inputs and outputs from an off-the-shelves simulator, We show that a neural network can learn a surrogate model effectively and efficiently and thus can be used as a surrogate simulation model. Moreover, we argue that the neural network model, although trained on some example terrains, is generally capable of simulating terrains of different sizes and spatial characteristics.