Session Abstracts:
Assistant Professor
Hong Kong University of Science and Technology
Surge-NF: Neural Fields inspired peak storm surge surrogate modeling with multi-task learning and positional encoding
Co-Authors: Wenjun Jiang (Hong Kong University of Science and Technology) and Xi Zhong (Hong Kong University of Science and Technology)
Abstract: Storm surges pose a significant threat to coastal communities, necessitating rapid and precise storm surge prediction methods for long-time risk assessment and emergency management. High-fidelity numerical models such as ADCIRC provide accurate storm surge simulations but are computationally expensive. Surrogate models have emerged as alternative to alleviate the computational burden by learning from available numerical datasets. However, existing surrogate models face challenges in capturing the highly non-stationary and non-linear patterns of storm surges, resulting in over-smoothed
response surfaces. Moreover, the dry-wet status of nearshore nodes has not been informatively considered in the training process.
This study proposes Surge-NF, a novel point-based surrogate model inspired by Neural Fields (NF) from computer graphics. Surge-NF introduces two key innovations. A positional encoding module is proposed to mitigate over-smoothing of high-frequency peak storm surge spatial dependencies. A multi-task learning framework is proposed to simultaneously learn and predict the dry-wet status and peak surge values, leveraging task dependencies to improve prediction accuracy and data efficiency. We evaluate Surge-NF on the NACCS database with comparison to state-of-the-art alternative surrogate models. Surge-NF consistently reduces RMSE/MAE by 50% and achieves 4-5 times computational cost gain over baselines, requiring only 50 training storms to produce accurate predictions. The complementary benefits of the positional encoding and multi-task learning modules are evident from the improved prediction capability with their combined use.
Professor
University of Notre Dame
Regional risk estimation using Gaussian Process metamodeling: application to storm-surge hazard
Co-Authors: WoongHee Jung (University of Notre Dame), Norberto Nadal-Caraballo (Army Corps of Engineers), Madison Yawn (Army Corps of Engineers), and Luke Aucoin (Army Corps of Engineers)
Abstract: The recent, very active hurricane seasons, as well as emerging concerns related to the future effects of sea level rise, storm intensification, and increased hurricane occurrence rate on coastal areas, make the prediction of storm-flood hazard a key priority when discussing coastal community resilience. To address this priority, researchers have placed substantial efforts in developing improved high-fidelity, numerical models to predict the storm-surge for a given storm event. For promoting computational efficiency when utilizing these models within hazard estimation applications, surrogate modeling techniques have emerged as a popular strategy. The accuracy of such techniques in this context has been examined so far using cross-validation (CV) techniques or by testing their performance for a (very) small number of historical storms. This presentation investigates this topic within a different setting, examining the resultant regional storm surge hazard maps, specifically using Gaussian Process (GP) as surrogate model choice. This is accomplished by examining the hazard products (hazard maps or curves) obtained by GP implementations, as well as the hazard products established through alternative, simplified Monte Carlo approaches. Examining this accuracy fills in an important knowledge gap and provides an answer to the question “what are the benefits in coastal hazard estimation by using surrogate models?”, improving at the same time trustworthiness of the associated results within the context of coastal risk quantification. The selection of the storm ensemble supporting the GP development is also examined, and it is shown that an adaptive implementation provides distinct advantages.
Graduate Student Researcher
University of Notre Dame
Computationally Efficient Regional Hazard Response Prediction Leveraging Sparse Graph Neural Networks
Co-Authors: Alexandros Taflanidis (University of Notre Dame)
Abstract: Regional risk assessment has been receiving increased attention within the natural hazards engineering community. Key barriers for this assessment is how to accommodate (i) the use of computationally expensive numerical models or (ii) applications with large number of outputs (geographical locations or assets) of interest. Data-driven predictive modeling has emerged as a critical tool for addressing the first barrier. Among the various approaches, Graph Neural Networks (GNNs) have gained significant attention since not only do they demonstrate high accuracy, but also they have the ability to capture sophisticated dependencies across the responses (outputs). The GNN implementation in this context encounters challenges from the second barrier discussed above, as they cannot easily scale to high-dimensional outputs. To address this issue, this study introduces a novel and scalable GNN framework, integrating sparsity in the graph connectivity. Central to the framework is the construction of graph structure using a cross-correlation technique that incorporates dependencies both within a region and across multiple hazard observations. Additionally, the framework employs dimensionality reduction to further extend the scalability to even higher dimensional outputs. The framework is demonstrated for storm surge response emulation using datasets of high-fidelity numerical simulations developed during regional flood studies. This topic is of increased relevance for improving the resilience of coastal communities. The proposed GNN formulation is shown to efficiently incorporate data dependencies while providing deeper insights into regional hazard dynamics.
Graduate Student Researcher
University of Texas at Austin
Efficient Coastal Flood-Inundation Mapping Using Machine Learning-Enhanced Static Model
Co-Authors: Jun-Whan Lee (The University of Texas at Austin) and Wonhyun Lee (The University of Texas at Austin)
Abstract: The increasing frequency and intensity of tropical storms, driven by climate change, have escalated the vulnerability of coastal regions, endangering lives, infrastructure, and ecosystems. To address these risks, rapid and accurate hazard simulation models are needed to produce coastal flood inundation maps using real-time data. Traditional high-fidelity hydrodynamic models, while precise, are computationally intensive and less suited for real-time applications. In contrast, static models are computationally efficient, but tend to overestimate flood extents by applying uniform or spatially varying water levels across landscapes and comparing them to topographical features and bathymetry. To overcome this limitation, we improved a static coastal flood model that utilizes water level data from coastal gauges and high-resolution digital elevation models with Bayesian optimization (machine learning algorithm) into the calibration process. This calibration integrates land cover data and high-water marks from historical storm events to generate reduction factors used to adjust flood extents post-calibration. The results demonstrate that machine learning-enhanced model significantly increased the precision of flood extent predictions by 8% compared to pre-calibration results, as validated against a physics-based numerical model (Super-Fast Inundation of Coasts; SFINCS) using Hurricane Ike (2008) as a case study. This approach maintains the computational speed necessary for real-time applications. Moreover, it offers a robust tool for hazard simulation and emergency response planning in coastal regions.
Graduate Student Researcher
University of Notre Dame
Efficient real-time probabilistic storm surge risk assessment through adaptive multi-fidelity Monte Carlo
Co-Authors: Alexandros Taflanidis (University of Notre Dame), Aikaterini Kyprioti (University of Oklahoma), and Jize Zhang (Hong Kong University of Science and Technology)
Abstract: Real-time probabilistic predictions for anticipated storm surges provide valuable information for emergency preparedness/response decisions during landfalling storms. These predictions are made through an uncertainty quantification process that involves: generating an ensemble of storm scenarios based on the nominal storm advisory and the anticipated forecast errors; performing high-fidelity numerical simulations to obtain surge predictions for each storm scenario; estimating surge statistics of interest by assembling the simulation results. This process is repeated whenever the storm advisory is updated. The need to make predictions in real-time drives the demand for high computational efficiency. This work investigates an adaptive Multi-Fidelity Monte Carlo (MFMC) framework to address the demand. As a lower-fidelity model within the MFMC setup, a surrogate model is adaptively developed based on high-fidelity simulations from the current or past storm advisories. MFMC leverages the correlation between the high- and low-fidelity models to establish unbiased predictions with high statistical accuracy and significant computational savings. To facilitate the development of the low-fidelity model using a small number of high-fidelity simulations, dimensionality reduction is introduced for both the input and output of the model. To accommodate the use of high-fidelity simulations from the current advisory in the development, leave-one-out model predictions are utilized in the MFMC estimation. Finally, the challenge of establishing MFMC predictions for a large number of quantities of interest (QoIs) corresponding to surges at different geographical locations is discussed. These QoIs may advocate conflicting decisions for the optimal MFMC implementation, and an efficient search for a compromise solution is introduced.
Postdoctoral Scholar
University of California, Santa Cruz
Introducing coastal ecosystems in flood risk models for valuing nature-based solutions for coastal resilience and climate adaptation
Co-Authors: Pelayo Menéndez (University of California, Santa Cruz), Chris Lowrie (University of California, Santa Cruz), David Gutierrez Barcelo (University of California, Santa Cruz), Rae Taylor-Burns (University of California, Santa Cruz), Brook Constantz (University of California, Santa Cruz), Camila Gaido (University of California, Santa Cruz), Borja Reguero (University of California, Santa Cruz), and Michael W. Beck (University of California, Santa Cruz)
Abstract: Coastal habitats such as mangroves and coral reefs can reduce flood risk, naturally adapt to climate change, and benefit the community economically. However, most global hazard simulations and damage models fail to include the effects of habitats. We developed a methodology to evaluate these habitats' flood risk reduction benefits at different spatial scales. We applied the methodology to estimate the current and future coastal floods and these habitats' ability to reduce the risks across reef and mangrove coastlines. Coastal hydrodynamic model SFINCS is used to simulate 10-, 25-, 50-, and 100-year return period events’ flooding for the current and future scenarios. The current waves and sea levels are taken from historical reanalysis of wave climate and hindcast, and future waves and sea level projections from the SSP5-8.5 emission scenario. We considered one SLR scenario based on median IPCC projections. We assessed four ecosystem scenarios: current reefs and mangroves, no mangroves, no reefs, and no reefs or mangroves. Then, we calculated the economic impact using the global urban footprint, depth damage functions, and national macroeconomic values of the assets exposed to coastal flooding. Results suggest that the current coastal flood risks in the Dominican Republic are about $30 million, which would increase by an additional $100 and $155 million without the protection of existing coral reefs and mangroves in current conditions and the future, respectively. Our findings demonstrate the critical role that nature-based solutions play in coastal adaptation and support restoring the habitats within adaptation, hazard mitigation, and disaster recovery decisions.