Spherical DYffusion Simulates 100 Years of Climate in 25 Hours

Spherical DYffusion is a generative AI model that projects 100 years of global climate patterns in 25 hours, a task that takes conventional models weeks of supercomputer time. Developed by computer scientists at the University of California San Diego and the Allen Institute for AI in Seattle, it is roughly 25 times faster than the state of the art while remaining nearly as accurate. That matters because climate simulations are the foundation of every major climate decision, from emissions policy to city infrastructure planning, and their enormous cost currently limits how many scenarios scientists can explore.

The model, presented at the NeurIPS conference in Vancouver in December 2024, emerged from a collaboration between the lab of Rose Yu, a faculty member in the UC San Diego Department of Computer Science and Engineering, and researchers Brian Henn, Oliver Watt-Meyer and Christopher S. Bretherton at the Allen Institute for AI (Ai2). The project originated in an internship that Yu's PhD student Salva Ruhling Cachay completed at Ai2.

Why Climate Simulations Are So Expensive

Traditional climate models divide the planet into a three-dimensional grid and solve the physics equations governing the atmosphere, ocean, land and ice at every cell, step by step, across decades or centuries of simulated time. These simulations are among the most computationally demanding workloads in science. They run on the world's fastest supercomputers and can take weeks to complete a single long run.

That cost has a direct policy consequence. Because each simulation is so expensive, scientists and policymakers can only run simulations for a limited amount of time and consider a limited number of scenarios. Questions like how a specific regional adaptation strategy plays out under several different emissions pathways, or how a rare compound extreme event behaves, often go unanswered simply because the compute budget is not there.

Diffusion Models Meet the Atmosphere

The researchers' key insight was that diffusion models, the generative AI technique behind popular image generators, could be adapted for ensemble climate projection. A diffusion model learns to transform random noise into realistic data by reversing a gradual noising process. Applied to climate, it can learn the statistics of atmospheric patterns and generate plausible future states of the planet.

Conventional diffusion models are slow because they require many sequential denoising steps. The team addressed this by pairing the diffusion approach with a Spherical Neural Operator, a neural network architecture designed to work with data distributed on a sphere, which is exactly what global atmospheric data is. The resulting model starts with knowledge of climate patterns and applies a series of learned transformations to predict future patterns, producing large ensembles of climate trajectories far more efficiently than a standard diffusion model could.

As the researchers note, it may be possible to generate just as realistic and accurate predictions with conventional diffusion models, but not with the same speed. The spherical architecture is what makes the difference, and it also means the model runs on GPU clusters in a research lab rather than requiring a national supercomputing facility.

What an Emulator Changes

Spherical DYffusion is an emulator, meaning it learns from the output of a conventional physics based climate model rather than replacing the underlying science. The physics model remains the source of truth, and the emulator learns to reproduce its behavior at a fraction of the cost. This distinction matters for trust: the AI system is anchored to simulations built on decades of climate physics rather than inventing weather from scratch.

In its published form, the model emulates the atmosphere, which Rose Yu describes as one of the most important elements in a climate model. The researchers acknowledge limitations and aim to include more elements in future iterations. Their next steps include simulating how the atmosphere responds to CO2, a critical capability for projecting warming under different emissions scenarios.

Why Speed Is a Public Good

Faster and cheaper climate simulation shifts who gets to ask questions. A national weather agency with a modest GPU budget could run large ensembles of regional projections. University researchers could test hypotheses that today require a grant proposal and a supercomputer queue. Cities could evaluate local adaptation plans against many possible futures instead of a handful. The researchers write that data driven deep learning models are on the verge of transforming global weather and climate modeling, and Spherical DYffusion is a concrete demonstration of that shift.

The work sits alongside a broader wave of AI weather and climate systems, including DeepMind's GenCast and GraphCast, the European Centre for Medium Range Weather Forecasts' AI based forecasting experiments, and NVIDIA's Earth 2 digital twin. Together these efforts point toward a future where the cost of asking the planet a question about its future drops by an order of magnitude, and the number of people who can ask grows along with it.