Image credit: Caltech Science and Engineering Institute
Much of the buzz surrounding artificial intelligence and climate change centers on how AI can be used to monitor weather events or track greenhouse gas emissions. Those are excellent uses of AI, but researchers want to dig deeper. To better understand the future impact of climate change, researchers are applying AI and machine learning to large-scale climate models, hoping to more accurately predict how changes in atmospheric conditions affect weather on the ground.
Climate modeling software is hampered by the sheer size of what must be assessed. A single pixel in a large-scale climate map can represent 50 or more miles, making weather predictions literally fuzzy. Weather maps divide the Earth into a grid, and each square requires hundreds of calculations to predict weather patterns, as climate changes are determined by location, time, and neighboring squares. The more detailed the model, the more computing power is needed. The entire process can bog down even the most powerful supercomputer.
A Different Type of Cloud Computing
To reduce the calculation time and improve the accuracy of weather predictions, a team of researchers at the University of California, Irvine, in collaboration with Columbia University and the Ludwig Maximilian University of Munich, is applying AI to cloud models. Knowing how clouds behave can help researchers predict climate change. Clouds have a significant impact on weather: they trap, transport, and release water and absorb and reflect the sun’s rays and heat. They are much smaller than a square on the grid, however, so they require a lot of computing power.
The UCI researchers trained deep learning systems to understand cloud behavior. The “Cloud Brain” has been trained on short-term runs of high-resolution cloud models and then tested to ensure it produces the same climate simulations as tried-and-true methods. Training the algorithms took about three months. The predictions were accurate and about 20 times faster than traditional processes. New algorithms are integrating the effects of mass and energy into the predictions. The research team hopes to integrate the small-scale cloud-specific AI into a large-scale global climate model.
The deep learning algorithms are specific to its initial training, so the Cloud Brain doesn’t adapt well to new scenarios, such as a warmer climate or different terrain. Using short-term models to train the AI also could affect the predictions when requesting a long-term forecast. Regardless, researchers are confident this use of AI will improve the current grid-based models.
The research team is hopeful that a similar machine learning approach could be applied to other large-scale global models. The technology could shed light on how large eddies alter ocean currents and how mountain ranges impact rainfall.
On Cloud Nine
With a similar goal but a different approach is the Climate Modeling Alliance (CliMA), a collaboration between California Institute of Technology, NASA’s Jet Propulsion Laboratory, the Naval Postgraduate School, and MIT. CliMA researchers are building from the ground up a new Earth modeling system that integrates advances in computer architectures and artificial intelligence and is based on Julia, a dynamic programming language developed at MIT. It is designed to automatically learn from multiple data sources and integrate observational and simulation data, particularly those around eddies and clouds.
CliMA envisions the new Earth modeling system will scale to high-resolution models more accurately to predict rainfall, heat waves, and droughts. With refined data, climatologists will be able to better forecast extreme weather conditions, flood risks, and crop yields, and they can better predict how climate change will impact our future on Earth.
- Find out more about CliMA.