Diffusion models enhance Earth turbulence inference from sparse data

Researchers have explored four generative diffusion modeling approaches for super-resolution and inference of quasi-geostrophic turbulence on a beta-plane, using coarse, sparse, and gappy Earth system observations. Two guided methods, SDEdit and Diffusion Posterior Sampling (DPS), adapt pre-trained unconditional models, while two conditional models require retraining with paired data. The study found that SDEdit produced unphysical results, and DPS generated smoothed features. Conditional models, though requiring more effort, reconstructed fine-scale features, maintained cycle-consistency with observations, and accurately predicted turbulence statistics. AI

IMPACT Demonstrates potential for diffusion models to improve resolution and inference of complex physical systems from limited observational data.

RANK_REASON Academic paper detailing novel application of diffusion models to a scientific problem. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Anantha Narayanan Suresh Babu
• beta plane
• diffusion posterior sampling (DPS)
• Earth
• SDEdit

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →