INVESTIGATION OF DIMENSIONALITY REDUCTION TECHNIQUES IN THE SURROGATE MODEL DEVELOPMENT FOR SPATIOTEMPORAL STORM SURGE PREDICTIONS

Abstract

Coastal communities face unique challenges due to their increased vulnerability and exposure to natural hazards, with humans and their assets being at a close proximity to the ocean and to a variety of severe weather phenomena such as tropical cyclones. Storm surge, which is the abnormal rise in seawater level due to storm winds, can lead to catastrophic flooding, erosion and infrastructure damage, putting lives and properties at risk. With the frequency of extreme weather phenomena like tropical cyclones increasing due to climate change the need to be able to predict their impact in coastal areas is increasing. Accurate predictions of storm surge, as a tropical cyclone is approaching, are crucial for informing evacuation planning efforts, allowing communities to have enough lead time to prepare and respond effectively. Through robust predictions which provides timely, location-specific information, such vulnerable regions resilience can be improved by saving lives and reducing property damage, thereby aiding recovery efforts. In this effort to obtain reliable predictions, advances in computational and numerical models, have enabled the development of physics-based models, that can predict storm surge based on complex fluid dynamics equations and atmospheric conditions. These models are the closest approximation possible to the physical phenomena that contribute to the experienced surge, providing valuable and detailed insights for large regions of interest. Unfortunately, due to their inherent complexity, they are computationally intensive, requiring large memory and number of processors, with the time for each analysis necessitating a couple of days in a state-of-the-art supercomputing facility. These limitations prevent the use of such accurate and detailed models during real-time risk assessment efforts, where a specific tropical cyclone has been formed and is hours away from landfall. Especially these settings, it is of crucial importance to be able to produce time-series storm surge estimates of high spatial resolution fast and accurately so that preparedness and recovery efforts can be guided appropriately and utilized efficiently, maximizing their impact to the extent possible. In an effort to overcome these limitations, surrogate models have emerged as a versatile alternative, that offer a fairly simple, tractable, mathematical approximation of a computationally expensive model, allowing the production of fast and accurate surge predictions, provided that they are properly calibrated on available datasets for the region of interest. Such efficiency without the significant loss of any accuracy makes them a promising alternative in real-time emergency planning and response when resources and time to respond are limited. This requires the development of surrogate models that will be able to handle large geospatial regions providing detailed predictions in the most computationally efficient way to allow for enough lead time for the evacuation managers to develop their plans, exploring at the same time a large number of potential stormevolving scenarios Pursuing model simplicity and computational efficiency, past efforts have explored the development of surrogate models of different mathematical complexity, that leverage the underlying spatial correlation that the grid locations have within the domain of interest. This allowed for linear and nonlinear dimensionality reduction techniques like Principal Component Analysis (PCA) and AutoEncoder (AE) architectures to generate latent space representations, that can be used as the reduced size input during the surrogate model development and deployment phases. After predictions are made in that reduced, latent space, the actual surge predictions are compiled through an inverse mapping. So far, such investigations have not thoroughly examined nonlinear dimensionality reduction techniques for large geospatial domains where time-series predictions are warranted, especially for locations that are onshore where highly nonlinear surge behavior may be recorded during different tropical cyclone events. This thesis advances these efforts by performing a detailed investigation on the development of a nonlinear inverse mapping (utilizing the versatile family of autoencoders) designed to capture non-linear relationships for onshore nodes tackling challenges related to providing time-series storm surge predictions over a large number of regions. The latent space will be used as input in a Gaussian Process surrogate model to provide storm surge time-series predictions on a storm suite that the models have not been trained upon. This development seeks to support high-resolution forecasting, making it suitable for real-time emergency response as well as for long-term planning and infrastructure design for any region of interest, given the existence of a well-calibrated region-specific surrogate model.