Accelerating climate modeling with generative AI
The algorithms behind generative AI tools like DallE, combined with physics-based data, can be used to develop better ways to model Earth’s climate. Computer scientists in Seattle and San Diego have now used this combination to create a model that can predict climate patterns for more than 100 years 25 times faster than existing technology.
Specifically, the model, called “Spherical DYffusion,” can predict 100-year climate patterns in 25 hours, whereas for other models, this simulation would take weeks. Additionally, existing state-of-the-art models require running on supercomputers. The model can be run on a research lab’s GPU cluster.
“Data-driven deep learning models are on the verge of transforming global weather and climate models,” wrote researchers at UC San Diego and the Allen Institute for Artificial Intelligence.
The research team will present their work at the 2024 NeurIPS conference in Vancouver, Canada, December 9-15.
Due to their complexity, climate simulations are currently very expensive to generate. As a result, scientists and policymakers only have a limited time to run simulations and can only consider a limited number of scenarios.
One of the researchers’ key insights is that generative artificial intelligence models, such as diffusion models, can be used for overall climate predictions. They combined this with the spherical neural operator, a neural network model designed to process spherical data.
The resulting model starts with knowledge of climate patterns and then applies a series of transformations based on the learned data to predict future patterns.
“One of the main advantages of our model is its greater efficiency compared with traditional diffusion models (DMs),” the researchers wrote. “Using traditional diffusion models may produce equally realistic and accurate predictions, but not as quickly.”
In addition to running much faster than state-of-the-art techniques, the model is nearly as accurate and computationally inexpensive.
The researchers hope to overcome some of the model’s limitations in the next iteration, such as including more elements in the simulation. Next steps include simulating the atmosphere’s response to carbon dioxide2.
“We simulated the atmosphere, which is one of the most important elements in climate models,” said Rose Yu, a faculty member in the Department of Computer Science and Engineering at UC San Diego and one of the paper’s senior authors.
This work originated from an internship of Dr. Yu. Student Salva Ruhling Cachay of the Allen Institute for Artificial Intelligence (Ai2) conducted the research.
2024-12-02 20:01:54