Assessing physical realism in diffusion models for tropical cyclones

Tropical cyclones (TCs) are among the most destructive natural hazards worldwide. While several decades of satellite and reanalysis products now provide relatively large observational datasets of TCs, these datasets remain small by modern deep-learning standards and, crucially, are extremely imbalanced and do not sufficiently cover the tails of the distribution, with Category 5 cyclones being several orders of magnitude rarer than tropical storms. This severe data scarcity and imbalance poses fundamental limitations for supervised learning approaches to tasks such as intensity estimation, rapid intensification forecasting, or impact modeling, where performance on extremes is often the primary objective.

In this context, generative artificial intelligence offers a promising alternative. Diffusion models, in particular, have recently demonstrated state-of-the-art performance in modeling complex, high-dimensional data distributions. By learning the full probability distribution of TC-related fields rather than a single conditional mapping, diffusion models have the potential to generate physically plausible samples across the entire intensity spectrum, including rare but high-impact extremes. However, most existing applications of diffusion models—both within and outside the geosciences—are evaluated using perceptual or distributional metrics originally developed for natural images, such as visual inspection or feature-space distances. These metrics are poorly aligned with the physical constraints and scientific objectives that govern atmospheric phenomena, and may obscure important deficiencies in dynamical or thermodynamical realism.

Here, we present a diffusion-based generative framework for tropical cyclone spatial fields and propose a comprehensive evaluation strategy grounded in physically meaningful diagnostics. Rather than relying on perception-oriented scores, we assess generated samples using a suite of metrics designed to capture key aspects of TC structure and behavior, including radial symmetry, intensity–structure relationships, spatial gradients, and consistency with known climatological distributions across intensity classes. This allows us to directly interrogate whether the model reproduces physically coherent storm morphologies, particularly in the poorly sampled tails of the distribution. Beyond evaluation, we also explore multiple strategies for embedding physical realism directly into the model design. Together, these results highlight both the opportunities and the limitations of diffusion models as scientific tools for tropical cyclone research, and provide a framework for using generative AI not merely as a data-augmentation device, but as a principled instrument for studying rare and extreme atmospheric phenomena.