Mammoths: A Vital Nutrient for Early American Societies

Scientists have uncovered the first direct evidence that ancient Americans relied primarily on mammoth and other large animals for food. Their research sheds new light on both the rapid expansion of humans throughout the Americas and the extinction of large ice age mammals. Scientists have uncovered the first direct evidence that ancient Americans relied primarily
HomeTechnologyHarnessing Generative AI to Transform Climate Modeling

Harnessing Generative AI to Transform Climate Modeling

 

The algorithms that power generative AI tools like DallE, when paired with physics-based data, can enhance our abilities to model the Earth’s climate. Researchers in Seattle and San Diego have utilized this combination to develop a model that can predict climate trends over a century at a speed 25 times faster than current leading methods.

Specifically, the model, known as Spherical DYffusion, is capable of forecasting 100 years of climate trends in just 25 hours—a task that would typically take weeks for other models. Furthermore, while existing cutting-edge models depend on supercomputers for processing, this new model can operate on GPU clusters found in research facilities.

The researchers from the University of California San Diego and the Allen Institute for AI state, “Data-driven deep learning models are on the verge of transforming global weather and climate modeling.”

The research team will share their findings at the NeurIPS conference 2024, scheduled for December 9 to 15 in Vancouver, Canada.

Generating climate simulations is presently a costly endeavor due to their intricate nature. Consequently, scientists and policymakers are limited to running simulations for shorter durations and examining only a narrow range of scenarios.

A significant insight from the researchers was that generative AI models, like diffusion models, could facilitate ensemble climate projections. By integrating this approach with a Spherical Neural Operator—a neural network tailored to handle spherical data—they achieved their results.

The model starts with an understanding of existing climate trends and subsequently employs a series of transformations based on acquired data to forecast future patterns.

“Our model’s primary advantage over traditional diffusion models (DM) is its remarkable efficiency. While it’s feasible to produce equally realistic and precise forecasts using standard DMs, they do not match our model’s speed,” the researchers explained.

Beyond its speed, the model also boasts a high degree of accuracy without incurring the high computational costs associated with traditional methods.

However, the model does have certain limitations which the researchers intend to address in future updates, such as incorporating additional elements into their simulations. Future work includes modeling the atmosphere’s response to increased CO2 levels.

“We replicated the behavior of the atmosphere, which is a crucial factor in any climate model,” mentioned Rose Yu, a faculty member in UC San Diego’s Department of Computer Science and Engineering and one of the senior authors of the study.

This research originated from an internship completed by one of Yu’s Ph.D. students, Salva Ruhling Cachay, at the Allen Institute for AI (Ai2).