Estimation of extreme tropical cyclone risk using AI-weather models

With its global footprint, AXA is exposed to multiple natural hazards across the globe. Assessing the frequency and intensity of these events, especially unobserved extremes, is crucial to monitor, mitigate and adapt to the risk they pose.

Tropical cyclones are one of the most scrutinized natural risks by global (re)insurers. Curated observational records date back to the mid-1800s, with increased reliability from the satellite era onwards (post 1970). They are a global risk, with temporal and spatial dependencies between tropical basins. The extreme damage they cause has been at the root of the development of Natural Catastrophe (NATCAT) modelling capabilities by specialized modelling firms, brokers, and (re)insurers.

However, as exposure is increasing and climate is changing, especially in tropical cyclone prone coastal areas globally, the need for robust and accurate estimates of the frequency and intensity of adverse impacts from tropical cyclones is expanding. Observational tropical cyclones datasets like IBTrACS are too short to obtain reliable statistics on rarest and most impactful events.Fine resolution numerical weather models are too computationally expensive to run on extended periods of time.

AI-based weather models running on GPU-accelerated compute infrastructure provide the necessary speedup while maintaining physical accuracy, enabling the generation of thousands of synthetic tropical cyclone seasons. Using NVIDIA's Earth-2 platform, we build a pipeline to produce hundreds of downscaled large ensemble predictions.

This study investigates the potential of these downscaled runs to generate large sets of tropical cyclones physically consistent in space, time and intensity, yielding robust estimates of their impact probability, especially for the rarest events.