MIT's climemu Generates Climate Projections in 1 Second on a Single GPU
A score-based diffusion model on a spherical mesh generates monthly climate fields in about one second on a mid-range GPU, matching ESM output statistics.
ProTakeaways
- MIT team releases score-based diffusion emulator producing monthly ESM outputs in ~1 second on a single mid-range GPU
- Model uses HEALPix spherical mesh UNet with only ~10M parameters and 50MB file size
- Emulates temperature, precipitation, relative humidity, and wind speed conditioned on GMST via pattern scaling
- Errors stay smaller than ESM internal variability across three CMIP6 models on held-out scenarios
- Precipitation in strong seasonal-shift regions (like ITCZ zones) remains a known weakness
- Code, pretrained weights, and PyPI package available at github.com/shahineb/climemu
Climate scientists have a scaling problem that has nothing to do with model size. Every few years, policy makers propose new emissions pathways, but the Earth System Models (ESMs) needed to project their consequences take years to run on supercomputers. A new paper in the Journal of Advances in Modeling Earth Systems shows that a lightweight score-based diffusion model can stand in for these simulators, producing statistically faithful monthly climate fields in roughly a second on a single GPU.
The work, from a group at MIT's Department of Earth, Atmospheric and Planetary Sciences, targets the specific niche of impact-relevant emulation: providing the temperature, precipitation, relative humidity, and wind speed fields that downstream impact models (crop yields, water balance, energy demand) actually consume.
The gap diffusion is filling
Computational cost has kept ESMs from keeping pace with updates to emissions scenarios. The CMIP7 scenarios are expected imminently, but the corresponding ESM simulations will not be available for another few years. That leaves impact analysts working from outdated projections when they most need current ones.
Existing emulators mostly took one of two shortcuts. Many focus on one or two variables at a time, often making parametric assumptions like a lognormal distribution for precipitation or a Gaussian one for temperature. Others stitch together slices of pre-run ESM archives, which preserves multivariate consistency but cannot generate genuinely new realizations. Neither approach scales gracefully when you need dozens of correlated variables at native ESM resolution to study compound risks, since the number of correlations grows quadratically with the number of variables and spatial resolution.