PhD Student University at Buffalo
Rapid Spatio-Temporal Prediction of Hurricane-Induced Flooding Via Knowledge-Enhanced Graph Neural Network
Co-Author: Teng Wu (University at Buffalo)
Abstract: Severe hurricane-induced flooding has caused significant loss of life and property in many coastal cities, highlighting the need for reliable and rapid flood prediction tools to assess resilience enhancement strategies. While purely data-driven models are computationally less demanding than hydrodynamic models, their training requires large datasets, and their predictions can be physically implausible. Integrating domain knowledge in the framework of machine learning can potentially address both these challenges. In this study, equation-based domain knowledge is leveraged during model training by integrating the mass conservation law into the loss function. The obtained loss function is discretized to enforce the conservation of predicted water volume at each cell within the computational domain. Since mass conservation is enforced directly on each cell of the same mesh used by the hydrodynamic model for data labeling, the water mass leaving each cell is balanced by the fluxes entering neighboring cells. In addition, the derivative terms of the loss function in a discretized format are computed from the predicted output field rather than by employing automatic differentiation. Consequently, the training efficiency is significantly improved, and the extreme sensitivity of predictions to gradient values is alleviated. Topological domain knowledge is also utilized by transforming the unstructured mesh into a graph. The obtained graph neural network (GNN) maps storm surge time series at the shoreline (with detailed topological features) to spatio-temporal inundation depth and water velocity. The proposed knowledge-enhanced GNN is able to rapidly provide physically consistent and locally accurate spatio-temporal flood predictions during a hurricane event.
Assistant Professor The Hong Kong University of Science and Technology
Spatially interpretable dimensionality reduction for storm surge analysis: a unified treatment to overcome missing data, model bias and high dimensionality
Co-Authors: Xi Zhong (The Hong Kong University of Science and Technology)
Abstract: High-fidelity hydrodynamic models and sensor networks generate massive datasets essential for coastal risk assessment; however, these resources are frequently compromised by incompleteness (dry nodes), observation sparsity, and high dimensionality. Traditional reduction techniques, such as Principal Component Analysis (PCA), struggle to accommodate missing data and produce abstract global modes that lack physical interpretability. This study presents a novel data-driven framework based on Non-negative Matrix Factorization (NMF) to address these operational challenges in ocean engineering. Unlike global methods, NMF decomposes surge fields into additive, spatially localized patterns that align with physical sub- regions, allowing for the native handling of missing values without pre-imputation. We demonstrate the framework's utility through three engineering applications using the NACCS storm surge database. In imputation, we recover missing data in dry nodes with high accuracy; in bias correction, we significantly reduce model error by projecting sparse high-fidelity gauge data; in surrogate modeling, NMF aids to predict peak surges with accuracy comparable to PCA but with interpretable reduced dimensions. The results establish NMF as a reduction technique, but more beyond as a practical spatial- aware tool for coastal modelers to manage incomplete data and enhance the reliability of storm surge risk assessments.
Research Fellow University of Massachusetts Amherst
Global Peak Storm Surge Prediction
Co-Authors: Prateek Arora (University of California Berkley) - Presenting, Jinpai Zhao (University of Texas at Austin) and Eirik Valseth (Norwegian University of Life Sciences)
Abstract: We present a novel synthetic ADCIRC tropical cyclone simulation database with over 15,000 storms and global coverage. This database is intended to support surrogate modeling efforts and is unique in its size and global scope. We also present a surrogate model of storm surge baked on our new database that has global applicability. Our surrogate model leverages neural network architectures from computer vision and focuses on predicting maximum surge profiles. We demonstrate accurate prediction results across six different ocean basins: the North Atlantic, North Indian, South Indian, West Pacific, East Pacific, and South Pacific basins. Interestingly, we discover that surrogate models trained on the entire global dataset outperform those trained on basin-specific data, indicating that our model is able to learn location-independent storm surge dynamics. Our global model represents a step forward in storm surge emulation capability from traditional surrogate models which are tightly coupled to a small geographic region.
Associate Professor University of South Florida
Storm Surge and Coastal Inundation Nowcasts/Forecasts During Hurricanes Helene and Milton
Co-Authors: Haibo Xu (University of South Florida), Kaili Qiao (University of South Florida), Sebin John (University of South Florida), Robert H. Weisberg (University of South Florida), Jing Chen (University of South Florida), Lianyuan Zheng (University of South Florida), Sherryl Gilbert( University of South Florida), Steven A. Murawski (University of South Florida), Gary T. Mitchum (University of South Florida), and Thomas K. Frazer (University of South Florida)
Abstract: A daily automated coastal water level (storm surge) nowcast/forecast guidance system has been developed by the USF Ocean Circulation Lab based on the West Florida Coastal Ocean Model (WFCOM) and the very high-resolution Tampa Bay Coastal Ocean Model (TBCOM). Both models are configured to perform realistic simulations of ocean circulation and water levels which are then combined with tide gauge observations to provide 3-day hindcasts and 3.5-day forecasts of coastal water level along the West Florida coast (http://ocgweb.marine.usf.edu/Models/SeaLevel/). The experimental product was maintained during the approach and passage of Hurricanes Helene and Milton, and provided critical storm surge forecasts to a broad suite of stakeholders including the public. The system successfully predicted the water level set-up and set-down along the west Florida coast three days in advance of each hurricane, with improved forecasts realized each day prior to landfall. The TBCOM-inundation forecast system was also activated during Hurricane Helene. This modeling system extends its dense grid onto the land, facilitating simulation of inundation and flooding associated with storm surge in coastal areas. During Hurricane Helene, areas of severe inundation were identified along the coastal periphery of Tampa Bay and forecasts were accessible two days in advance of landfall.
Postdoctoral Scholar Oregon State University
A Building-Aware Computational Workflow for Urban Storm Surge Modeling with Dynamic Structural Collapse
Co-Authors: Daniel Cox (Oregon State University) and André Barbosa (Oregon State University)
Abstract: Computational workflows for coastal flood hazard assessment support community-scale risk analysis, yet most operational surge simulations rely on bare-earth digital elevation models (DEMs) that omit explicit representation of buildings and their evolving interaction with flow, such as successive collapses. This contribution presents a building-aware storm surge modeling workflow applied to Hurricane Ian (2022) at Estero Island, Florida, designed to advance computational methods for resolving urban-scale overland flow modification and cascading failure processes.
Using Delft3D-FM on an unstructured mesh with 2–4 m resolution in the built environment, building footprints derived from GIS data are incorporated as hydraulic structures with parameterized resistance and height. This formulation enables explicit simulation of flow redirection, velocity modulation, and shielding effects while remaining computationally tractable for large urban domains. To extend beyond static representations, a dynamic collapse mechanism is implemented using real-time control (RTC), allowing building properties to change during runtime based on prescribed failure parameters.
Simulation results across no-building, intact-building, and dynamic-collapse scenarios show that while regional surge elevations are similar, local velocity fields and inferred shear zones vary substantially. Peak velocities increase by up to 1 m/s in urban corridors and newly opened pathways following structural failure, shifting hazard hotspots toward previously sheltered areas. These differences have direct implications for scour potential, debris transport, and emergency accessibility that are not captured by depth-based products alone.
This workflow complements SimCenter efforts focused on uncertainty quantification and loss modeling by providing physically informed, building-aware hazard fields suitable for downstream integration in community-scale risk assessment frameworks.
Professor University of Memphis
A Scalable Framework for Multi-Hazard Flood Damage and Resilience Assessment in Urban Coastal Communities
Co-Author: Maryam Karimi
Abstract: Urban coastal regions are increasingly vulnerable to extreme flooding driven by hurricanes, sea-level rise, and compound coastal hazards, resulting in escalating economic losses, infrastructure damage, and threats to public safety. This study presents a comprehensive and scalable framework for quantifying community-scale flood damage and structural resilience across multiple spatial scales. Large-scale coastal flooding scenarios were simulated and validated using semi-coupled storm surge and two-dimensional inundation models. Key hydrodynamic parameters, including flood depth and flow momentum, were integrated into a newly developed multidimensional flood-damage assessment model.
The proposed framework extends traditional depth–damage relationships by incorporating building-specific characteristics such as height, age, configuration, and construction materials. Structural resilience was quantified as a function of flood-induced damage, recovery time, and community preparedness. Application of the framework to a highly vulnerable coastal study area revealed flood depths ranging from 4.91 m to 8.06 m under multiple hurricane categories, resulting in property damage levels of 28.69%, 45.62%, and 92.13% across scenarios.
Results were analyzed using geospatial techniques at both the property (microscale) and census block group (macroscale) levels, demonstrating the framework’s adaptability and scalability. By explicitly linking hydrodynamic processes with structural and community-level attributes, this work provides actionable insights to support hazard mitigation planning and enhance resilience in urban coastal systems facing recurrent extreme flooding.