Improving Storm Surge Surrogate Modeling Through Advances in Integration of Spatial Correlation Information

Surrogate models (also known as emulators or metamodels) have emerged as powerful, data-driven predictive tools to support storm surge risk assessment applications. In this context, they are developed using a suite of synthetic storm simulations, leveraging existing databases created during regional flood studies. Once calibrated, these surrogate models can efficiently and accurately predict the expected storm surge for unobserved storms, and therefore can replace the original high-fidelity model (that was used to create the database) within various risk-assessment tasks. The most popular emulation approaches for this application are Gaussian Process metamodels (GPMs). The output approximated corresponds to the storm surge (maximum values or time-history evolution) for different geographic locations of interest, and the input to the features that can be used to describe each storm. The geographic locations, termed herein as nodes for simplicity, may correspond directly to the computational nodes of the underlying numerical simulation model or to some few, representative locations, typically referenced as save points (SP). Depending on the extent of the geographic domain, the number of such locations may vary from few thousands (for SP), up to more than one million (for computational nodes). Development of metamodels in this context needs to consider the spatial correlation between nodes. This thesis formally considers the use of Gaussian-Process based geointerpolation (GPI) for this objective. From a modeling perspective, the key contribution of this thesis the development and implementation (in a GPI setting) of an adaptive covariance tapering methodology to accommodate the large number of nodes described by the resultant covariance matrix. The adaptive covariance tapering yields sparse covariance matrices that accommodate GPI implementation even for datasets with large numbers of nodes. Applications of the resulting GPI is considered for both pure geo-interpolation (e.g., data imputation) and establishing storm surge predictions at unobserved nodal locations. The databases used in this thesis consist of synthetic storm simulations developed via ADCIRC, a high-performance, high-fidelity numerical simulation tool used to model storm surge, tides, and coastal circulation. Applications to the United States North Atlantic and the Gulf coasts are examined.