Spherical Diffusion Model Projects 100 Years of Climate in 25 Hours Using AI
The UC San Diego and Allen Institute for AI teams combined generative AI with physics-based climate data to create the Spherical DYffusion model. This model simulates a century of climate patterns in merely 25 hours, a task that traditionally takes weeks or months. It leverages diffusion techniques adapted for spherical data to maintain spatial accuracy over long-term projections.
This breakthrough illustrates how hybrid AI-physics models can drastically accelerate complex scientific simulations without sacrificing fidelity. For AI practitioners, it emphasizes the value of domain integration—melding data-driven models with established physical laws to enhance prediction speed and reliability. This approach could transform workflows in environmental science and beyond.
UC San Diego researchers and the Allen Institute for AI developed and validated the Spherical DYffusion model, enabling faster climate projections that could inform policy and adaptive strategies more responsively.
Step 1: Access the Spherical DYffusion codebase or similar diffusion-based climate models from UCSD’s repository (check https://github.com/ucsd-ai-climate). Step 2: Prepare your climate dataset ensuring compatibility with spherical coordinates (e.g., NetCDF files). Step 3: Run the model training and inference scripts to generate accelerated long-term climate projections, observing significant reductions in simulation time compared to traditional methods.