The award for designing the foundation model includes 650,000 hours of computational time on advanced supercomputing systems, which is worth around $5 million in funding.
Dr. Sujit Roy, a computer scientist in the Lab for Applied Science (LAS) at The University of Alabama in Huntsville (UAH), a part of the University of Alabama System, led the effort to secure a major research resource grant funded by the U.S. National Science Foundation (NSF) in partnership with dozens of federal agencies and non-governmental organizations. Dr. Roy collaborated with UAH researchers in the Center for Space Plasma and Aeronomic Research (CSPAR) to submit a proposal for access to computational resources needed to develop and train the first artificial intelligence (AI) foundation model for heliophysics.
Supported by NASA’s Interagency Implementation and Advanced Concepts Team, Roy has worked with UAH’s Earth Systems Science Center since July 2022, serving as the lead AI researcher for the foundation models team. He is particularly focused on developing foundation models for a variety of scientific domains. These large-scale models are pre-trained on vast amounts of data and serve as a starting point, or “foundation”, that can be fine-tuned for more specific data analysis tasks. After a single FM is constructed, it can be utilized in a variety of applications, which ultimately reduces computational burden and cost.
Roy says that the heliophysics and space weather FM he is building will enable new research approaches and capabilities in these fields. “This model aims to not only enhance space weather forecasting, but also push the boundaries of what is currently possible in predicting and studying solar activities and their effects on space weather.” For example, an FM for heliophysics and space weather will make it possible to more accurately predict solar flares and coronal mass ejections – phenomena that can significantly affect satellite operations, communication systems, and power grids.
While there are many benefits to developing domain-specific FMs, they are complex and require significant computation, storage, and human resources to build and train. Roy intends to use high-resolution datasets to train the FM. Fortunately, NASA’s Solar Dynamics Observatory (SDO) spacecraft has collected 14 years worth of quality solar observation data that is unparalleled in its comprehensive coverage and extended duration.
Color-enhanced image of a solar prominence eruption captured by NASA’s Solar Dynamics Observatory on March 30, 2010.
Credit: NASA Solar Dynamics Observatory team
Training an FM at this scale would be a lengthy, years-long process without the assistance of supercomputers. Therefore, receiving the National Artificial Intelligence Research Resource (NAIRR) award was a critical first step towards building the model that Roy and his collaborators envision. Currently operating as a pilot program, NAIRR supports U.S. researchers with access to federally-funded computation, data, software, model and training resources to advance AI research. Roy and his team learned in July that their FM proposal was accepted by NAIRR and that they would be granted 650,000 GPU hours of computational time on 32 NVIDIA DGX systems optimized for machine learning. In dollars, this award is worth close to $5,000,000.
“This project is particularly suited for support from NAIRR,” says Roy. “We anticipate NAIRR facilitating advanced research and collaboration opportunities and helping impact broader scientific communities. We also plan to utilize NAIRR to publish the large-scale, machine learning-ready datasets and to train the community to use and fine-tune the foundation model.”
To learn more about the NAIRR pilot program, visit the NAIRR website.