Understanding and predicting compound coastal flooding remains one of the most challenging problems in hydrologic and coastal sciences. Flooding in coastal environments usually emerges from the interaction of tides, storm surge, waves, river discharge, precipitation, groundwater, and evolving sea levels, often acting across multiple spatial and temporal scales (Green et al). These complexities have motivated remarkable advances in numerical modeling, observational systems, machine learning, and conceptual frameworks. Yet, despite substantial progress within each modeling paradigm, research has largely evolved along parallel trajectories rather than through systematic integration. As highlighted by recent perspectives, continued reliance on individual modeling approaches beyond their intended strengths risks limiting scientific progress and preventing the community from fully leveraging complementary capabilities (Nazari et al). The next generation of coastal flood prediction therefore lies not in replacing existing methods, but in strategically combining them.
Recent studies have emphasized that hybrid modeling should be viewed as a distinct scientific framework rather than simply coupling multiple models together. By integrating computational, statistical, observational, and data-driven approaches, hybrid systems capitalize on the unique advantages of each while mitigating their individual limitations. Emerging classifications, including sequential, feedback, and ensemble hybrid architectures, provide a useful foundation for comparing different integration strategies and identifying where information should flow between component models (Radfar et al). This perspective moves the field beyond isolated model development toward flexible modeling ecosystems that are both scientifically rigorous and operationally practical.
The importance of such integration is evident in recent efforts to develop comprehensive coastal forecasting systems. Large-scale modeling frameworks now routinely combine multiple hydrodynamic, wave, hydrologic, and atmospheric models to simulate the interactions among coastal, fluvial, and pluvial flooding processes within unified computational environments (i.e.Ā Green et al. (2026),Ā Zhao et al, andĀ Sun et al). These integrated systems represent a major step toward physically consistent prediction of compound hazards across diverse coastal landscapes. Their strength lies in faithfully resolving complex physical interactions that cannot be represented by individual models alone. However, this level of physical realism comes at a substantial computational cost, limiting their application for probabilistic analyses, long-term climate assessments, uncertainty quantification, ensemble forecasting, and real-time decision support where thousands of simulations may be required (Jafarzadegan et al).
At the other end of the spectrum, recent years have witnessed rapid advances in data-driven techniques for flood prediction. By leveraging large observational and simulation datasets, these approaches have demonstrated remarkable capability in learning complex nonlinear relationships while producing forecasts with exceptional computational efficiency. However, purely data-driven models remain fundamentally constrained by the information they can provide (i.e. hazard level at gauged locations) and the information contained within their training datasets. Their predictive skill often deteriorates when confronted with unprecedented events, evolving climate regimes, or hydrodynamic conditions that lie outside the historical record, limiting their reliability for extrapolation and high-impact decision making (Irish,Ā Ragno,Ā Bahmanpour). Equally important, many machine learning models provide limited physical interpretability, making it difficult to diagnose the governing processes underlying their predictions or establish confidence for applications involving rare extremes (Bao et al). Recent comparative assessments further demonstrate that model architecture strongly influences generalization, with sequence-aware designs such as CNNāLSTM substantially outperforming conventional neural networks for unseen flood events, while purely spatial architectures often struggle to reproduce the temporal evolution of compound flooding (Fen et al, andĀ Daramola et al). These findings reinforce that statistical accuracy alone is insufficient and that robust flood prediction requires learning strategies capable of representing both the spatial and temporal dynamics of physical flood processes (Wang et al., andĀ Yarveysi et al).
Hybrid methodologies have consequently emerged as a promising pathway for overcoming these complementary limitations by integrating the strengths of process-based understanding with the efficiency of machine learning. Rather than treating physics-based and data-driven models as competing paradigms, hybrid frameworks increasingly embed physical knowledge directly within learning architectures through transfer learning, physics-informed loss functions, conservation constraints, and adaptive coupling strategies. Recent studies demonstrate that transfer learning can substantially reduce computational demands while improving model portability across diverse coastal environments, provided that physical understanding is incorporated into the learning framework (Daramola et al). More recently, physics-informed neural networks have shown that explicitly enforcing governing equations, including mass and momentum conservation, substantially improves predictive skill, physical consistency, and robustness under data-scarce conditions while maintaining inference speeds orders of magnitude faster than conventional numerical models (Yang et al., andĀ Radfar et al). Hybrid statistical-dynamical frameworks are transforming regional and long-term flood hazard assessments. Recent developments combine stochastic generators of compound flood drivers with process-based hydrodynamic simulations and surrogate machine learning models capable of reproducing complex flood responses at only a fraction of the computational expense (seeĀ Wang et al.,Ā Emmanouli et al., andĀ Ricondo et al.Ā for example). Related reduced-order frameworks replace repeated high-resolution simulations with statistical emulators trained on carefully selected physics-based model outputs, enabling rapid mapping of coastal, fluvial, and pluvial flooding over large ensembles of environmental conditions. These approaches dramatically expand the range of questions that can be addressed, from centennial-scale hazard assessments and climate adaptation planning to probabilistic risk analysis and operational forecasting.Ā
Taken together, these developments suggest that the future of compound coastal flood modeling will not be defined by increasingly sophisticated individual models, but by thoughtfully designed hybrid frameworks that integrate complementary sources of information across observations, numerical simulations, statistical inference, and machine learning. Such systems have the potential to achieve an effective balance between physical realism, computational efficiency, predictive accuracy, and interpretability. As coastal hazards intensify under a changing climate and the demand for actionable flood information continues to grow, hybrid modeling provides a compelling roadmap toward scalable, transferable, and scientifically robust solutions capable of supporting both fundamental research and real-world decision making.
