Reviewed by Lexie CornerDec 3 2024
By combining physics-based data with algorithms powered by generative AI tools like DALL-E, improved methods for modeling the Earth's climate are being developed. Computer scientists in San Diego and Seattle are using this approach to create a model that predicts climate patterns over a century, performing 25 times faster than current methods.
Image Credit: Fahroni/Shutterstock.com
The model, Spherical DYffusion, can simulate 100 years of climate patterns in just 25 hours, whereas other models take weeks. Unlike existing cutting-edge models, which require supercomputers, this model can run on GPU clusters in a research lab.
“Data-driven deep learning models are on the verge of transforming global weather and climate modeling,” said researchers from the University of California San Diego and the Allen Institute for AI.
The research team will report their findings at the NeurIPS 2024 conference, which will be held in Vancouver, Canada, from December 9th to 15th, 2024.
Climate models are currently expensive to produce due to their complexity, which limits the ability of scientists and policymakers to run simulations over extended periods or explore a wide range of scenarios.
A key finding of this study is that generative AI models, such as diffusion models, can be used to generate ensemble climate predictions. The team combined this approach with a Spherical Neural Operator, a neural network designed to process data on a sphere.
The resulting model starts with existing climate trend data and uses a series of transformations based on learned information to predict future climate patterns.
The researchers added, “One of the main advantages over a conventional diffusion model (DM) is that our model is much more efficient. It may be possible to generate just as realistic and accurate predictions with conventional DMs but not with such speed.”
In addition to being significantly faster than current models, the new model is nearly as accurate while being much less computationally demanding.
However, the model does have limitations, which the researchers aim to address in future versions, such as incorporating more variables into their simulations. The next steps will focus on simulating the environmental impact of CO2.
We emulated the atmosphere, which is one of the most important elements in a climate model.
Rose Yu, Study Senior Author and Faculty Member, Department of Computer Science and Engineering, University of California San Diego
The study is based on an internship at the Allen Institute for AI (Ai2) completed by Salva Ruhling Cachay, one of Yu’s Ph.D. students.