A collaboration involving NASA and IBM Research has led to the development of a new artificial intelligence (AI) foundation model for weather and climate: Prithvi-weather-climate (Prithvi is the Sanskrit name for Earth). The model is pre-trained on 40 years of weather and climate data from NASA's Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2), and fills a need to infuse AI and machine learning (ML) methods into weather and climate applications, such as storm tracking, forecasting, and historical analysis.
In keeping with NASA's open science policies, Prithvi-weather-climate will be openly available. The model and the code will be released later in 2024 through Hugging Face, a public repository for open-source ML models.
The Role of Foundation Models
Using AI to sift through data to find solutions requires not only massive amounts of data, but also large amounts of time. As noted by IBM Research, the next stage in AI model development is to create models pre-trained on a broad set of unlabeled data that can be used as the foundation for different tasks that require minimal fine-tuning. These are called foundation models, or FMs.
FMs are the basis for enabling AI and ML systems to ingest large amounts of data and generate results based on associations among the data. They serve as a baseline from which scientists can develop a diverse set of applications that can result in powerful and efficient solutions. Once an FM is created, it can be trained on a small amount of data to perform a specific task.
But creating and pre-training FMs still requires lots of data. When it comes to addressing the need for massive amounts openly available Earth science data, NASA's Earth Science Data Systems (ESDS) Program is a logical source. The more than 100 petabytes (PB) of data the program distributes openly and without restriction is the secret sauce that helps make open-source Earth science-based FM development possible. This combination of open NASA Earth science data and IBM Research's state-of-the-art computational power led to the groundbreaking work using NASA Harmonized Landsat and Sentinel-2 (HLS) data to create the Prithvi Geospatial FM, the first open-source geospatial FM, in 2023. Prithvi-weather-climate builds on this achievement.
"Foundation models offer amazing prospects for expanding the use of NASA’s vast archive of Earth observations," says NASA Earth Data Officer Katie Baynes. "The Prithvi-weather-climate model holds promise to advance our understanding of atmospheric dynamics and developing new applications. We're excited to see how the community can leverage this work to enhance resilience to climate and weather-related hazards."
Creating Prithvi-weather-climate
Work on Prithvi-weather-climate began in September 2023 with a workshop at NASA's Marshall Space Flight Center in Huntsville, AL. Marshall is the home of NASA's Interagency Implementation and Advanced Concepts Team (IMPACT), a NASA ESDS element charged with expanding the use of NASA Earth observation data through innovation, partnerships, and technology—including the application of AI to these data.
"This model is part of our overall strategy to openly and collaboratively develop a family of AI foundation models to support NASA's science mission goals," says IMPACT Manager Dr. Rahul Ramachandran. "These models will augment our capabilities to draw insights from our vast archives of Earth observations."
Joining the IMPACT and IBM Research teams in developing Prithvi-weather-climate were participants from NASA Headquarters, NASA's Global Modeling and Assimilation Office (GMAO), NASA's Center for Climate Simulation (NCCS), the NASA Advanced Supercomputing (NAS) Facility, Oak Ridge National Laboratory (ORNL), and NVIDIA Corporation. Several universities engaged in various aspects of AI or large-scale computing and weather/climate science participated as well, including the University of Alabama in Huntsville, Colorado State University, and Stanford University.
The focus of the Marshall workshop was to plan the next six to eight months of work necessary to develop and pre-train the model. It was decided that the FM would contain parameters such as wind speed and direction, air temperature, specific humidity, cloud mass variables, and longwave and shortwave radiation variables. To be valuable to the broader science community, the team agreed that the focus should not be on forecasting; rather, the FM should enable many different types of downstream science applications.
MERRA-2
The foundation of Prithvi-weather-climate is 40 years of MERRA-2 data. MERRA-2 is the first long-term global reanalysis to assimilate space-based observations of aerosols and represent their interactions with other physical processes in the climate system. These data are available through NASA's Earthdata Search. MERRA-2 was created by NASA's GMAO to replace and enhance the original MERRA and to sustain GMAO's commitment to having an ongoing near real-time climate analysis.
"With the Prithvi-weather-climate FM, NASA and IBM have led the creation of a unique AI representation of all knowledge available in 40 years' worth of MERRA-2 data," says Dr. Juan Bernabé-Moreno, director of IBM Research Europe and IBM’s accelerated discovery lead for climate and sustainability. "The IBM-NASA collaboration highlights how open-source technologies are essential to advancing crucial research into areas such as climate change. By merging IBM's foundation model technology with NASA's deep expertise and specialized climate datasets, we've developed flexible, reusable AI systems for broader use."
Applications to Science and Society
The Prithvi-weather-climate FM has broad applications for both science and society.
From a scientific and research standpoint, the model has been fine-tuned to increase the resolution of long-term climate models by a factor of 12x, a process known as "downscaling." Using an AI model in this context avoids the high costs associated with the conventional approach using high performance computing (HPC). The FM also improves the use of AI for better representation of small-scale physical processes in numerical weather and climate models. Through the insertion of tokens in the model at wind turbine locations, Prithvi-weather-climate can generate targeted forecasts using hyper-localized, asset-specific observations, further improving the accuracy of short to medium-range forecasts.
The application of AI to weather and climate data also can lead to improvements in public safety. Uses being developed by the research team for Prithvi-weather-climate include more precise hurricane track and intensity forecasts along with better seasonal precipitation forecasting. As the model continues to be trained, future applications include the detection and prediction of severe weather patterns, more detailed wildfire behavior forecasts, finer turbulence detection and prediction, urban heatwave prediction, and improved solar radiation assessment.
"Our ambition is to accelerate and advance the impact of NASA's Earth science to meet this moment of changing climate for the benefit of all humankind," says Dr. Karen St. Germain, director of NASA's Earth Science Division. "The [NASA/IBM Research] foundation model for weather and climate will enable this Earth Science to Action strategy."
Meet the Team
Along with the participants noted in the Creating Prithvi-weather-climate section, the development of the FM was accomplished by a diverse team with deep experience representing varied aspects of AI and ML.
| Prithvi-weather-climate Element | Element Leads |
|---|---|
| Model Development: NASA, University of Alabama in Huntsville (UAH), and ORNL | Dr. Sujit Roy, UAH Dr. Arlindo da Silva, NASA GMAO Dr. Valentine Anantharaj, ORNL |
| Model Development: IBM | Dr. Johannes Schmude Dr. Johannes Jakubik Dr. Anne Jones Dr. Julian Kuehnert Dr. S. Karthik Mukkavilli Dr. Daniel Salles Civitarese Shraddha Singh Dr. Daniela Szwarcman Dr. Will Trojak |
| Science Expertise | Dr. Arlindo da Silva, NASA GMAO |
| Use Case Development | Dr. Christopher Phillips, UAH Dr. Rajat Shinde, UAH Ankur Kumar, UAH Vishal Gaur, UAH Dr. Amy Lin, UAH Dr. Aristeidis Tsaris, ORNL Dr. Simon Pfreundschuh, Colorado State University Dr. Aman Gupta, Stanford University |
| Infrastructure | Mike Little, NASA's Goddard Space Flight Center Dr. Mark Carroll, NASA's Goddard Space Flight Center Shubha Ranjan, NASA's Ames Research Center, NAS Bill Thigpen, NASA's Ames Research Center, NAS |
| Leadership/Technical Oversight | Dr. Rahul Ramachandran, NASA IMPACT Dr. Manil Maskey, NASA IMPACT Dr. Tsengdar Lee, NASA Headquarters Dr. Juan Bernabé-Moreno, IBM |
Learn More
- Building an AI Foundation Model for Weather and Climate
- NASA, IBM Research Release New AI Foundation Model for Weather, Climate Forecasts
- IBM and NASA are building an AI foundation model for weather and climate
- Blog: Environmental analysis made easier with IBM’s Geospatial Studio
- Mukkavilli, S.K., et al. (2023). AI Foundation Models for Weather and Climate: Applications, Design, and Implementation. Cornell University arXiv. doi:10.48550/arXiv.2309.10808
- Gelaro, R., et al. (2017). The Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2). Journal of Climate, 30(14): 5419-5414. doi:10.1175%2FJCLI-D-16-0758.1
