Combining a stochastic climate emulator with surrogate models of dynamic coastal simulators to drive coastal flood impacts research

Naval Base Coronado, San Diego, CA
Submitted by:
Peter Ruggiero
Oregon State University
Project URLs:

Project Description

The ability to understand and forecast coastal flooding events and associated impacts are critical societal needs within the context of projected sea level rise and climate change. State of the art practice in estimating extreme events uses the observational record of water levels. However, such projections often do not account for the individual contributions of processes resulting in compound TWL events, nor do they account for time-dependent probabilities due to seasonal, interannual, and long-term oscillations within the climate system. Further, numerical models that combine processes such as tides, storm surge, wave propagation and wave runup are typically too computationally expensive to dynamically simulate the full parameter space of both present and future oceanographic, atmospheric, and hydrologic conditions that will constructively compound to cause both extreme event and nuisance flooding. To address these issues, we exploit a range of data mining and machine learning techniques by integrating the stochastic climate emulator TESLA (Time-varying Emulator for Short- and Long-term Analysis of coastal flooding, Anderson et al., 2019) with surrogate models of process based dynamical simulators. The efficient hybrid statistical-dynamical framework is applied here to probabilistically explore coastal flood impacts along both the oceanside and bayside of San Diego, CA. TESLA is a stochastic climate emulator in which each relevant physical process associated with coastal flooding is simulated through an auto-logistic regression accounting for the interannual variability of ENSO, seasonality, intra-seasonal propagation of the MJO, and daily variability of local sea level pressure and sea surface temperature fields. TESLA is capable of simulating thousands of synthetic years of a representative climate to produce potential combinations of forcing variables contributing to flooding that have not necessarily been seen in the historical record. We combine output from TESLA with computationally efficient surrogate models developed to emulate the relevant output from high fidelity process-based hydrodynamic simulators (Delft3D and XBeach components of the Coastal Storm Modeling System, CoSMoS, O’Neill et al., 2018) to dynamically downscale the forcing conditions produced by TESLA at local to regional scale. On the oceanside, we develop surrogate models (e.g., gaussian process regression, random forest) of wave runup, overtopping, and inundation extent at over 20 individual cross-shore transects. In San Diego Bay we make use of spatially variable water level emulation via Principal Component Analysis rather than creating individual surrogate models for individual grid nodes, which would be computationally prohibitive.

Key Successes

Developed new techniques including TESLA; Developed depth/consequence relationships at installations that include downtime/days as well as capital expenditures/dollars


takes a lot of time, dialog to establish the sensitivity relationships; Difficult to handle massive amounts of stochastic model output
Any question regarding this posting, email