Abstract. Growing urban development, combined with the influence of El Niño and climate change, has increased the threat of large unprecedented floods induced by extreme precipitation in populated areas near mountain regions of South America. High-fidelity numerical models with physically based formulations can now predict inundations with a substantial level of detail for these regions, incorporating the complex morphology, and copying with insufficient data and the uncertainty posed by the variability of sediment concentrations. These simulations, however, typically have large computational costs, especially if there are multiple scenarios to deal with the uncertainty associated with weather forecast and unknown conditions. In this investigation we develop a surrogate model or meta-model to provide a rapid response flood prediction to extreme hydrometeorological events. Storms are characterized with a small set of parameters, and a high-fidelity model is used to create a database of flood propagation under different conditions. We use kriging to perform an interpolation and regression on the parameter space that characterize real events, efficiently approximating the flow depths in the urban area. This is the first application of a surrogate model in the Andes region. It represents a powerful tool to improve the prediction of flood hazards in real time, employing low computational resources. Thus, future advancements can focus on using and improving these models to develop early warning systems that help decision makers, managers, and city planners in mountain regions.
Training Machine Learning Surrogate Models From a High‐Fidelity Physics‐Based Model: Application for Real‐Time Street‐Scale Flood Prediction in an Urban Coastal Community
