There’s a lot of hype about generative AI, both pro and con. Researchers at the University of California, San Diego and the Allen Institute for Artificial Intelligence (Ai2) are on the pro side, demonstrating that it can have valuable global impact. They have developed a generative AI climate prediction model they call Spherical DYffusion, which is fast and agile enough to be used as a tool not just by scientists, but by anyone whose decisions are affected by climate trends.
In their recently published paper online, the researchers explained the difference between climate modeling and weather forecasting. “Climate models are foundational tools used to understand how the Earth system evolves over long time periods and how it may change as a response to possible greenhouse gas emission scenarios … There are fundamental differences between weather and climate modeling. Climate refers to the average weather over long periods of time. While weather forecasting focuses on short time scales in the order of days or weeks, climate modeling simulates longer periods of decades to centuries.”
“What people have usually been using for climate modeling are physics-based climate models, which solve equations about the atmosphere and oceans and so on,” said Salva Ruhling Cachay, one of the Ph.D. students of UCSD Professor Rose Yu. “These models are run on supercomputers to make predictions for 10s or even 100 years. But running such a physics-based model is very expensive in terms of time and compute and energy,” he added. Where it takes about six months to run a physics-based model, theirs was able to produce equally good results in about two weeks.
The reason the physics-based models take so long to produce results is that they rely on numerical calculations of the underlying differential equations. In contrast, the researchers call their model a “machine learning emulator.” It emulates a coarse
version of the United States’ primary operational global forecast model, the FV3GFS, which gives you good but very computationally expensive predictions of the climate. In order to create their emulator, the researchers run the FV3GFS a few times, for a 10-year projection. They then take that data and use it to train their machine learning-based emulator to serve as a substitute for the physics-based model.
Cachay also explained how Spherical DYffusion is better than other machine-learning models. “There are two key metrics for us that make it better than the previous machine learning models, which are first, the accuracy — how well we think it's projecting the future climate. And then the speed at which it does it. We think it finds a very good spot between speed and accuracy,” he said.
The secret sauce for their model is that it’s generative. Once it’s been trained on the data outputted by the FV3GFS, it’s able to generate multiple scenarios of the future, while still respecting the data that it’s been trained on. “The high-level idea of our generative model, is that it finds the pattern in the existing data in order to generate new data that did not exist in the training data,” said Yu. “It's similar to, for example, OpenAI DALL-E, which is trained on image data available on the Internet. Once the model is trained, you can generate new pictures that have not previously existed.”
In order to come up with predictions, they start with the initial state of atmospheric variables, and then add in forcing, or boundary, conditions such CO2 emissions, solar radiation, or ocean temperatures. “Although there's one model, there are multiple predictions because the model is stochastic. Every time you provide a deterministic input, it will give you a different output,” said Cachay. “We are interested in multiple predictions because the future is very uncertain. Because the whole Earth-system is chaotic, even if you change some temperature measurements very slightly, you might get a very different future.”
“These predictions could be used by government agencies to make informed policy decisions. Let's say they wanted to understand the global climate impact of electrical vehicle charging. They can use our model to quickly understand the impact of different policies on the future trajectories of the climate,” said Yu.
This article was written by Ed Brown, Associate Editor, SAE Media Group. For more information, visit here .
