AXA

For property and casualty insurers, accurately pricing catastrophe risk depends on robust statistical models of rare, high-impact events — yet historical hurricane records spanning only decades provide far too few extreme observations to build reliable loss distributions. Traditional statistical methods attempt to compensate by inflating historical datasets, but this approach lacks physical plausibility and cannot reproduce complex inter-event dynamics such as sequential storm occurrences within a single season or intensity interdependencies driven by atmospheric conditions. Physics-based numerical weather models are physically rigorous but demand massive compute resources — running a single year of simulation required one hour on 1,000 dual-socket CPU nodes — making the thousands of synthetic years needed for robust tail-risk statistics economically unattainable.

AXA partnered with NVIDIA to deploy the Earth-2 platform, leveraging the HENS-SFNO (Huge Ensemble — Spherical Fourier Neural Operator) AI weather model originally developed at UC Berkeley for extreme weather prediction. The implementation ran through NVIDIA's Earth2Studio Python library, enabling two distinct simulation workflows: targeted re-simulation of individual historic storms such as Hurricane Milton across multiple initialization dates (before cyclogenesis, at cyclogenesis, and approaching landfall), and generation of full synthetic hurricane seasons using six-week rollout chunks with two-week initialization intervals drawn from diverse historical years. Uncertainty was addressed through two complementary mechanisms — multiple independently trained model checkpoints to capture model-parameter uncertainty, and bred vector perturbation of initial conditions to capture atmospheric state uncertainty — producing calibrated, physically plausible ensemble members at scale.

The HENS-SFNO model delivered performance improvements that make large-scale catastrophe ensemble generation economically viable for the first time:

• Simulation speed: A one-year simulation completes in 12.8 minutes on a single GPU, versus 1 hour on 1,000 dual-socket CPU nodes using traditional IFS — an improvement of several orders of magnitude.
• Ensemble scale: A 1,024-member ensemble was generated for Hurricane Helene, initialized 39 hours before tropical storm categorization.
• Hurricane Milton: The model successfully reproduced the storm's anomalous Gulf of Mexico track and enabled analysis of counterfactual trajectories that could have struck different regions of Florida and the southern US coast.

Qualitatively, the approach unlocks full synthetic hurricane seasons, revealing frequency distributions and inter-event correlations that historical records alone cannot provide.

• AI weather models can reduce catastrophe simulation costs by orders of magnitude, but bias correction and spatial downscaling of near-surface wind speeds remain required before ensembles are production-ready for loss modeling.
• Simulating full synthetic seasons — not just individual storms — is essential for capturing frequency, intensity distribution, and inter-event correlations that drive tail-risk estimates.
• Using multiple trained model checkpoints alongside initial-condition perturbation addresses both model uncertainty and atmospheric state uncertainty, producing better-calibrated ensembles.
• Early initialization (39+ hours before tropical storm classification) is feasible with AI models, extending the actionable lead time for risk and emergency response workflows.
• Counterfactual scenario generation has dual value: improving statistical risk models and enabling emergency managers to evaluate alternative impact scenarios for mitigation planning.