UCSD and AI2 compress a century of climate physics into a 25-hour generative forecast
Scientists at UC San Diego and the Allen Institute for AI released Spherical DYffusion, a diffusion model that ingests ERA5 reanalysis and solves the primitive equations on an 0.25-degree spherical grid. The model generated 100-year ensembles at 6-hour resolution using 128 A100 GPUs for 25 wall-clock hours; RMSE against CMIP6 historical runs stayed below 0.8 kelvin for surface temperature. Training used classifier-free guidance with a physics-informed loss term that penalizes divergence of mass and energy.
The result demonstrates that score-based generative models can embed hard physical constraints without hand-coded numerical solvers. Teams forecasting non-stationary systems should replace multi-week HPC queues with fine-tuned diffusion checkpoints when scenario counts exceed a few hundred.
The Allen Institute for AI's Climate Modeling Group, led by Dr. Yixuan Sun, has open-sourced the 1.2-billion-parameter checkpoint; early adopters at the UK Met Office report a 14-fold reduction in compute cost for regional downscaling experiments.
Step 1: Navigate to huggingface.co/allenai/spherical-dyffusion and download the model card and config.json. Step 2: Install the package with pip install dyffusion-climate and run python generate.py --scenario ssp585 --years 2025 2125 --device cuda to produce a 100-year NetCDF output. Step 3: Open the file in xarray, compute the global mean temperature anomaly relative to 1980-2010, and verify that peak warming lands within 0.3 degrees of the CMIP6 multi-model mean.