IBM and NASA, along with contributions from Oak Ridge National Laboratory, have developed an innovative open-source AI model designed to address both short-term weather predictions and long-term climate projections. This flexible, scalable model offers a new way to handle challenges in meteorology and climatology, outperforming traditional models in various applications.
Key features of the model include:
- Targeted Forecasting: It can create localized forecasts based on specific observations.
- Severe Weather Detection: The model enhances the ability to predict severe weather events.
- Climate Simulation Improvement: It boosts the spatial resolution of global climate simulations and refines how physical processes are represented.
- Data Assimilation: In one test, the model accurately reconstructed global temperatures using only 5% of the original data, indicating its potential for handling data-limited situations.
The model was pre-trained on 40 years of Earth observation data from NASA's MERRA-2 (Modern-Era Retrospective Analysis for Research and Applications, Version 2). Its architecture allows for fine-tuning to various scales—global, regional, and local—making it adaptable to a range of studies.
Two key fine-tuned versions of the model are also available:
1. Climate and Weather Data Downscaling: This process infers high-resolution outputs from low-resolution inputs like temperature, precipitation, and wind data.
2. Gravity Wave Parameterisation: This version tackles the challenge of representing gravity waves in climate models, which have been traditionally underrepresented but are critical for understanding atmospheric processes like cloud formation and turbulence.
NASA’s Earth Science Division Director, Karen St Germain, highlighted the model’s potential for delivering actionable insights to communities and organizations, aiding in preparations for weather and climate-related events. IBM's Juan Bernabe-Moreno noted that the model goes beyond traditional AI models, which are often limited to single datasets and use cases, offering a more versatile tool for both scientific and industrial applications.
The AI model and its fine-tuned versions are available on Hugging Face, allowing broader access for weather and climate research.
In one experiment, the open-source AI model accurately reconstructed global surface temperatures from a random sample of only 5% of original data, suggesting a broader application to problems in data assimilation.
This model was pre-trained on 40 years of Earth observation data from NASA’s Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2).
We use cookies to ensure you get the best experience on our website. Read more...
