Generative AI neural network produces climate change model results nearly 25x faster than its predecessors

Researchers at the University of California San Diego have used artificial intelligence to model Earth's climate more faster and more complex than previous models at similar timescales. It may be a key step towards more real-time and applicable analyses of our continuing-changing climate.

AI
Deep learning is a machine learning process that utilizes neural networks to process data as a human brain would.


One of the primary challenges with environmental modeling is complexity. For weather modeling, the system (Earth) is complex and chaotic, so exorbitant computing power is needed for accurate modeling. The combined processing power of all U.S. weather forecast model supercomputers is 10,000 times faster than the average desktop computer.

For climate modeling, scientists are also undertaking Earth as a system. Although with even more complexity and across far-longer timescales into the future. This translates to time-plentiful time needed to run computer models. More time equals more money.

Artificial intelligence (AI) modeling is already beginning to revolutionize weather forecast modeling. Scientists have theorized that AI could also serve as an invaluable tool to increase the efficiency and decrease the cost of climate modeling. Now climate modelers at the University of California San Diego have demonstrated this with a generative AI model called Spherical DYffusion.

Neural networks

Neural networks are the key to Spherical DYffusion. Many AI models utilize machine learning through neural networks as a means to accomplish predictive modeling. Neural networks are so named as they are designed to mimic a human brain, allowing the processing of data in a way that parallels how humans think. A neural network is comprised of many different nodes often spread across different layers that can each be weighted differently.

Climate researchers at UC San Diego utilized the Spherical Neural Operator, an established type of neural network designed for application to a spherical data set such as Earth. With this design, a series of simulations can be run over time, with layers being weighted differently through successive runs. This mimics more of an ensemble-type forecast model, providing larger amounts of model output data to analyze.

Advantages of improved modeling

A primary advantage of this model over others is speed. Researchers noted that their method provides output 25-times faster than more traditional climate models over similar timescales. This could transform our understanding of how the climate is changing in a much more realistic timescale. Scientists would have the ability to run these models more frequently, providing initialization updates at a much finer time resolution scale.

Researchers noted that certain changing variables, like the concentration of greenhouse gases and aerosols, are not yet accommodated in this model. For more valuable insight, researchers need to incorporate model training that simulates these fluctuations.

These variables have a direct correlation with changing climate, so this is a key next step.

However, the reduced lag of these model runs due to this breakthrough is already a significant leap forward. More informed and timely analyses pertaining to policy-making decisions and climate-focused legislation can only help in combating the challenge of global climate change.