Reproduce “Physics-Informed Diffusion Model for Generating Synthetic Extreme Rare Weather Events Data” on OpenResearch

About this paper

Marawan Yakout, Tannistha Maiti, Monira Majhabeen, Tarry Singh

Researchers from the University of London and Deepkapha AI Lab developed a physics-informed diffusion model to generate synthetic multi-spectral satellite imagery of rare extreme weather events. This approach addresses severe data scarcity and class imbalance for rapidly intensifying tropical cyclones by producing physically consistent, class-specific wind-field patterns, achieving an average Log-Spectral Distance of 4.5 dB.

Key insights

• Physics-aware context conditioning effectively guides the generative process, enabling the model to produce distinct and physically plausible wind field patterns corresponding to various storm characteristics and lifecycle stages.
• The diffusion model successfully learns and reproduces complex non-linear fluid dynamics and spatial autocorrelation, generating coherent and diverse synthetic samples without mode collapse, a common issue in other generative models.
• The pre-generated noise strategy ensures consistent noise exposure across training epochs, which is crucial for maintaining pairing integrity and effectively learning from severely imbalanced datasets.

Summary by alphaXiv.