New AI model from IBM and NASA to transform weather forecasting capabilities

In a major breakthrough that promises to transform the landscape of meteorology and climate science, IBM (NYSE: IBM) and NASA have jointly released an open-source artificial intelligence foundation model designed for diverse weather and climate applications. This innovative model is now accessible to researchers, developers, and businesses on the Hugging Face platform, marking a significant step towards enhancing our understanding of weather patterns and climate change.

Advanced AI capabilities for weather forecasting

The newly launched foundation model, dubbed “Prithvi WxC: Foundation Model for Weather and Climate,” is distinguished by its advanced architecture and scalability, allowing it to address a wide array of challenges associated with short-term weather forecasting and long-term climate projections. Developed in collaboration with the Oak Ridge National Laboratory, this model surpasses traditional weather AI models by providing unparalleled flexibility in application. It enables targeted forecasts based on local data, the detection of severe weather patterns, and improved representation of physical processes in numerical models. Research published on arXiv reveals that the model can accurately reconstruct global surface temperatures from just five percent of original data, underscoring its potential for effective data assimilation.

Utilising decades of data for accurate predictions

The foundation model was pre-trained on an extensive dataset, comprising 40 years of Earth observation data from NASA’s Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2). This extensive training equips the model to handle various scales—global, regional, and local—making it particularly suited for diverse weather studies. It is available for download on Hugging Face, along with two fine-tuned versions tailored for specific applications.

One of these fine-tuned models focuses on climate and weather data downscaling, a crucial technique for generating high-resolution outputs from lower-resolution variables. This model can depict weather and climate data at resolutions up to 12 times finer than conventional approaches, generating localized forecasts and climate projections that are essential for accurate decision-making.

The second fine-tuned model addresses gravity wave parameterization, a critical factor in atmospheric processes. Gravity waves influence phenomena such as cloud formation and turbulence, yet traditional climate models often fail to adequately account for them, resulting in uncertainties in predictions. The new model aims to enhance our understanding of gravity wave generation, thereby improving the accuracy of numerical weather and climate simulations.

Expert insights on the significance of the model

Karen St. Germain, the director of the Earth Science Division at NASA’s Science Mission Directorate, emphasised the urgency of delivering actionable science that can assist communities in responding to the rapidly changing environment. She noted that this foundation model will equip users with tools for weather, seasonal, and climate projections, enhancing preparedness for future climate-related challenges.

Juan Bernabe-Moreno, Director of IBM Research Europe, explained that the model has been designed to move beyond the limitations of existing AI models, which often focus on fixed datasets and single use cases. He asserted that the model’s versatility enables it to operate on both global and local scales, making it invaluable for understanding complex meteorological phenomena like hurricanes and atmospheric rivers.

Arjun Shankar, director of the National Center for Computational Sciences at Oak Ridge National Laboratory, highlighted the collaboration’s objective to apply advanced computing and data to critical national issues. He underscored that the Prithvi weather and climate foundation model represents a significant advancement in computational science aimed at addressing pressing weather and climate challenges.

Collaborative efforts to enhance forecasting accuracy

IBM’s collaboration with Environment and Climate Change Canada (ECCC) seeks to further test the model’s flexibility in various forecasting scenarios. ECCC is currently exploring the potential of very short-term precipitation forecasting through a technique known as precipitation nowcasting, which integrates real-time radar data. The team is also experimenting with downscaling global model forecasts from a 15-kilometre resolution to even finer scales.

The release of this weather and climate foundation model is part of an ongoing collaboration between IBM Research and NASA to leverage AI technology for exploring our planet. This model complements the Prithvi family of AI foundation models, which includes the geospatial AI model launched last year, now the largest open-source geospatial AI model available on Hugging Face.

The foundation model and its variants can be accessed through the NASA-IBM Hugging Face page and the IBM Granite Hugging Face page, enabling widespread use across government, academia, and industry.