Near-future extreme sea level predictions from physics-informed deep networks

NLDL 2026 Conference Submission6 Authors

Keywords: LSTM, sea level, flood, CMIP projections, climate chang

TL;DR: Northern Europe will get more floods, but we can't say how much until the inputs improve

Abstract: Storm surges, high-frequency extreme sea level events driven by the atmosphere, lead to loss of life and crucial infrastructure. Their prediction is a requirement to build adequate coastal protection, yet by design the current generation of coupled climate models cannot reliably do so. They have however long been used to predict and project changes in the atmosphere. We here re-train a recently produced physics-informed Long Short Term Memory (LSTM) network that was designed to reproduce observed extreme sea level events around Northern Europe. We notably reduce its number of predictors to the six atmospheric variables that the network ranked as most important and its temporal resolution from hourly to 3-hourly, so that we can pass as predictors climate model-produced atmospheric variables. We compare the climate model-based and observed sea level over 1985-2014 and select the most accurate runs to quantify changes in the number of extreme sea level events in the near-future (2025-2054) and end of century (2070-2099). The network projects on average increases in the number of extreme events, but there is a large spread in the predictions. In general, it also projects a larger increase in the near-future than at the end of the century. We attribute this spread to the inconsistent changes in the drivers: larger wind speeds, shifting from more west-northerly in the near-future to more west-southerly at the end of the century, yet higher sea level pressure. These results show the feasibility of predicting changes in extreme events from physics-informed deep learning networks, but more reliable predictors, from a wider range of climate models, are needed before the predictions can converge.

Serve As Reviewer: ~Céline_Heuzé1

Submission Number: 6