NCAR will collaborate on new initiative to integrate AI with climate modeling
NSF announces new Center for Learning the Earth with Artificial Intelligence and Physics
Sep 10, 2021 - by Laura Snider
Sep 10, 2021 - by Laura Snider
The National Center for Atmospheric Research (NCAR) is a collaborator on a new $25 million initiative that will use artificial intelligence to improve traditional Earth system models with the goal of advancing climate research to better inform decision makers with more actionable information.
The Center for Learning the Earth with Artificial Intelligence and Physics (LEAP) is one of six new Science and Technology Centers announced by the National Science Foundation to work on transformative science that will broadly benefit society. LEAP will be led by Columbia University in collaboration with several other universities as well as NCAR and NASA’s Goddard Institute for Space Studies.
The goals of LEAP support NCAR’s Strategic Plan, which emphasizes the importance of actionable Earth system science.
“LEAP is a tremendous opportunity for a multidisciplinary team to explore the potential of using machine learning to improve our complex Earth system models, all for the long-term benefit of society,” said NCAR scientist David Lawrence, who is the NCAR lead on the project. “NCAR’s models have always been developed in collaboration with the community, and we’re excited to work with skilled data scientists to develop new and innovative ways to further advance our models.”
LEAP will focus its efforts on the NCAR-based Community Earth System Model. CESM is an incredibly sophisticated collection of component models that when connected can simulate atmosphere, ocean, land, sea ice, and ice sheet processes that interact with and influence each other, which is critical to accurately project how the climate will change in the future. The result is a model that produces a comprehensive and high-quality representation of the Earth system.
Despite this, CESM is still limited by its ability to represent certain complex physical processes in the Earth system that are difficult to simulate. Some of these processes, like the formation and evolution of clouds, happen at such a fine scale that the model cannot resolve them. (Global Earth system models are typically run at relatively low spatial resolution because they need to simulate decades or centuries of time and computing resources are limited.) Other processes, including land ecology, are so complicated that scientists struggle to identify equations that accurately capture what is happening in the real world.
In both cases, scientists have created simplified subcomponents — known as parameterizations — to approximate these physical processes in the model. A major goal of LEAP is to improve on these parameterizations with the help of machine learning, which can leverage the incredible wealth of Earth system observations and high-resolution model data that has become available.
By training the machine learning model on these data sets, and then collaborating with Earth system modelers to incorporate these subcomponents into CESM, the researchers expect to improve the accuracy and detail of the resulting simulations.
“Our goal is to harness data from observations and simulations to better represent the underlying physics, chemistry, and biology of Earth’s climate system,” said Galen McKinley, a professor of earth and environmental sciences at Columbia. “More accurate models will help give us a clearer vision of the future.”