How Initial Perturbations Affect AI-Driven Tropical Cyclone Predictions?

Accurate prediction of tropical cyclones (TCs) is critical for disaster mitigation, yet data-driven AI weather models often struggle with robustness and sensitivity to initial perturbations. This study conducts a comprehensive sensitivity analysis of an AI-based weather model (Pangu-Weather) to evaluate its resilience to geographic location and initial perturbations in TC positions and intensities, framed against seasonal average background conditions. Experiments focus on TCs in the tropical Atlantic and North Western Pacific basins, the feedback of the experiments could help us to understand the dynamic properties and physical reasons of AI-based weather models furtherly and the size of the initial perturbation for numerical design of future global ensemble systems.

Key findings reveal that TC track and intensity predictions are highly sensitive to initial perturbations in both the Atlantic and North Western Pacific basins, though larger perturbations are required for the Atlantic basin than for the North Western Pacific basin. It is confirmed that a data-driven AI model does follow up the dynamics and physics mostly. Notably, small symmetric perturbations (intensity: ±1–3 hPa from the mean surface pressure anomaly) do not generate meaningful error growth (spread) in the Atlantic basin—contrary to typical numerical model behavior—but they suffice for the North Western Pacific. The position perturbations (±50 km) also deviate from expected model responses for both basins. Meanwhile, initial spinup and geographic adjustment are discussed as well. These results may align with prior studies suggesting that AI models lack explicit physical constraints and dynamic conservation. Future work could explore targeted perturbations to further assess their impact on TC forecasts.