Reproduce “Probabilistic Emulation of a Global Climate Model with Spherical DYffusion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Salva Rühling Cachay, Brian Henn, Oliver Watt-Meyer, Christopher S. Bretherton, Rose Yu
Researchers from UC San Diego and the Allen Institute for AI present Spherical DYffusion, a conditional generative model for global climate emulation. This method efficiently produces probabilistic ensemble simulations by integrating dynamics-informed diffusion with spherical neural operators, reducing 10-year climate biases by over 50% compared to previous data-driven emulators while enabling 25x speed-ups over physics-based models.
Key insights
- Short-term weather forecasting skill does not directly correlate with the accuracy of long-term climate statistics; models optimized for short-term skill may accumulate systematic errors that distort long-term climate reproduction.
- Combining dynamics-informed diffusion with spherical-aware neural operators (SFNO) effectively enables efficient probabilistic climate emulation, overcoming limitations of both standard diffusion models and general-purpose architectures for spherical data.
- Probabilistic ensemble averaging substantially improves climate prediction accuracy, reducing biases by an additional 29.28% on average, highlighting the utility of generative approaches over purely deterministic ones for climate studies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.