IBM, in collaboration with the National Aeronautics and Space Administration (NASA) and Oak Ridge National Laboratory, has developed a new foundation model aimed at enhancing AI capabilities in climate and weather. Announced on Monday, this model has been specifically created to process climate modeling data to identify and forecast severe weather patterns, refine the spatial resolution of global climate simulations, and enhance the representation of physical processes in numerical weather and climate models. In addition to these primary applications, the foundation model is built to be scalable, catering to both short- and long-term weather forecasting. It can be downloaded from Hugging Face, an open-source AI code repository, and is part of the proprietary IBM-NASA Prithvi family of AI models, a series specifically for weather that derives its name from the Sanskrit word for “Earth. ” “Advancing NASA’s Earth science to serve humanity involves providing actionable science that is beneficial for individuals, organizations, and communities. The rapid changes we observe on our planet necessitate this approach to address the immediacy of the situation, ” remarked Karen St. Germain, director of NASA’s Earth Science Division, in a press release. “The NASA foundation model will enable us to create a tool for producing weather, seasonal, and climate projections that inform decision-making for preparedness, response, and mitigation strategies. ” Trained on four decades of Earth observation data from NASA's Modern-Era Retrospective analysis for Research and Applications, researchers indicate that the model features a specialized architecture capable of adapting to a wide range of global, regional, and local weather modeling applications. A study included in the documentation of the foundation model demonstrated its capacity to accurately reconstruct global surface temperatures using a dataset derived from only 5% of the original data. The model is specifically designed and trained for two key tasks.

The first is climate and weather downscaling, which involves deriving high-resolution outputs from low-resolution inputs. This aspect utilizes data such as temperature, weather forecasts, and surface winds — all characterized by varying resolutions. The model can enhance these data outputs by up to twelve times the original resolution, irrespective of input data quality. The second task is gravity wave parameterization, which aims to utilize current numerical climate models and data to illustrate the gravity waves in the atmosphere. Improved gravity wave generation is anticipated to enhance the accuracy of input numerical weather models by providing better estimations of gravity wave production and its effects. “This field has seen a surge of large AI models focusing on fixed datasets and single-use cases, primarily forecasting, ” said Juan Bernabe-Moreno, director of IBM Research Europe and IBM's Accelerated Discovery Lead for Climate and Sustainability. “We have designed our foundation model for weather and climate to transcend these limitations, allowing it to be adjusted to various inputs and applications. ” This initiative aligns with the Biden administration’s focus on harnessing innovation through public-private sector partnerships.