Constraining Estimates for South American Sea Level Extremes Using Uncertainty-Permitting Machine Learning

Sea level extremes, such as those arising from atmospheric storm surges, can economically devastate and endanger the lives of coastal populations. Predicting the risks that extreme sea level poses to coastal communities requires quantifying how various interacting factors—such as the local shelf geometry and shore-normal wind stress—impact extreme sea level. Here, we use a neural network trained on data from a high-resolution coupled climate model, GFDL CM2.6, to constrain estimates for South American coastal sea level extremes using only information about the local bathymetry and atmospheric forcing. The neural network is trained using a maximum-likelihood loss function, which quantifies an uncertainty range for the extremes given this minimal information. For a majority of the samples in the preindustrial control test set, our model predicts one-year return levels constrained to an uncertainty range with standard error less than 5cm. Furthermore, although our model is trained only on data from a preindustrial control simulation, it generalizes well to estimate extremes in a perturbation experiment, where CO2 concentrations are increased by 1% per year. This work shows how simple understanding of the bathymetry and atmospheric forcing variability might be sufficient to constrain estimates for sea level extremes—and illustrates how coastal communities can reevaluate changing vulnerabilities to extreme sea level under a changing climate.