EGU24-9533, updated on 08 Mar 2024
EGU General Assembly 2024
© Author(s) 2024. This work is distributed under
the Creative Commons Attribution 4.0 License.

Modeling inland flooding caused by tropical cyclones in the US using AI-based synthetic events

Nans Addor1,2, Natalie Lord1, Balaji Mani3, Thomas Loridan3, Naoki Mizukami4, Jannis Hoch1, and Malcolm Haylock1
Nans Addor et al.
  • 1Fathom, Bristol, UK (
  • 2University of Exeter, Exeter, UK
  • 3Reask, London, UK
  • 4National Center for Atmospheric Research, Boulder, USA

Tropical cyclones (TCs) are a key driver of flooding in the US. Here we present a modeling approach to simulate their associated inundation footprint under present and future climate and generate the hazard data necessary to run a CAT model. 

We developed an AI-based model called RainCyc that learns from the TC rainfall fields dynamically generated by the WRF model as well as from observations. RainCyc is orders of magnitudes faster than WRF, meaning that orders of magnitude more events can be simulated for the same computational cost. This is essential to capture the tail of the distribution, i.e., to generate synthetic events over a period longer than the longest return period of interest. Future boundary conditions for RainCyc are provided by the CESM2-LENS ensemble, which covers the 21st century under SSP370 levels of warming using 50 model realizations started from slightly perturbed initial conditions.

The rainfall fields produced by RainCyc are used to simulate inland flooding, i.e., pluvial and fluvial. The inundation footprint for each event is generated by sampling from flood hazard maps simulated by the LISFLOOD hydraulic model. The sampling for pluvial is informed by RainCyc precipitation, while for fluvial, it relies on hydrological simulations driven by the FUSE and mizuRoute models. FUSE is a frugal rainfall-runoff model that is run at 10km over a domain encompassing each event to generate its associated runoff. This runoff is then provided to the vector-based routing model mizuRoute to generate flow time series from which peak flow is extracted and used to sample fluvial hazard maps.

We present this modeling framework and test it for thousands of years of synthetic events under present and future climate. We benchmark the hydrological simulations for historical events using runs from other models, including GloFAS. We also test the ability of the framework to generate synthetic events spanning the intensities covered by hazard maps for a wide range of return periods.

How to cite: Addor, N., Lord, N., Mani, B., Loridan, T., Mizukami, N., Hoch, J., and Haylock, M.: Modeling inland flooding caused by tropical cyclones in the US using AI-based synthetic events, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-9533,, 2024.