CYCLONE: A superfast large-scale coastal storm surge model for Tropical Cyclones

We present CYCLONE, a deep learning framework based on Graph Convolutional Networks (GCNs) developed to predict tropical cyclone–induced coastal storm surge in the North Atlantic basin. The model generates a coastal storm surge peak map associated with a TC in less than one second.

CYCLONE was trained using tropical cyclone tracks well represented in ERA5 (Bourdin et al., 2022) and storm surge simulations generated with GSTM for the period 1980–2022. For the North Atlantic basin, this dataset includes a total of 247 tropical cyclones.

The core of CYCLONE relies on an architecture of Graph Convolutional Network layers. Each tropical cyclone is represented as an independent graph  instance, with  nodes corresponding to coastal stations and  edges defining the spatial connectivity of the coastline. The adjacency matrix with N coastal stations is fixed and shared across storms, allowing the model to learn spatially consistent patterns of surge propagation while remaining transferable across events and domains.

Training was performed using 80% of the available tropical cyclones. 170 tropical cyclones were used for training, while the remaining events did not generate significant storm surge and therefore did not contribute to the gradient computation. The remaining 20% of the storms (47 events) were used for validation.

CYCLONE is a tool capable of providing rapid, large-scale hazard assessments of tropical cyclones, especially in countries or with limited or no technical infrastructure. In this context, CYCLONE facilitates damage assessments and improves tropical cyclones response capabilities, which are essential for insurance, risk management and adaptation planning; key active areas of research in the context of climate change.