IBM and NASA have released Prithvi, one of the most sophisticated AI foundation models for climate and weather, on Hugging Face in conjunction with Climate Week 2024. Developed in collaboration with Oak Ridge National Laboratory, Prithvi extends beyond traditional AI forecasting, offering valuable insights for scientists, developers, and businesses.
Prithvi (Sanskrit name for Earth) was initially unveiled in May with the intent to be open-sourced later this year. Unlike existing models that primarily focus on forecasting, Prithvi is designed for a wide range of applications in both short-term weather predictions and long-term climate studies. Its unique architecture and training regime allow for enhanced precision, speed, and cost-effectiveness.
The model's release includes two specialized versions:
- A downscaling model that increases the resolution of weather and climate data up to 12 times, enabling more detailed local forecasts and projections.
- A gravity wave parameterization model, helping scientists better understand these atmospheric phenomena and their effects on processes like cloud formation.
Prithvi can be fine-tuned for various tasks with minimal additional data, demonstrating remarkable efficiency. In one experiment, it accurately reconstructed global surface temperatures using only 5% of the original data.
Sample of reconstruction as a “zero shot” skill of the foundation model. Scientists take an input, remove a percentage of the area and let the model fill that back in at the same time as the input. This example shows 2m surface temperature. 95% of data is removed and reconstructed.
The model was pre-trained on 40 years of Earth observation data from NASA's MERRA-2 dataset. Its architecture allows for adaptation to global, regional, and local scales, making it suitable for a wide range of weather and climate applications.
The release of Prithvi is part of a broader collaboration between IBM Research and NASA aimed at leveraging AI to explore and understand Earth’s geophysical processes. Last year, they introduced the largest open-source geospatial AI model on Hugging Face, which has since been utilized by governments and institutions to study disaster patterns, biodiversity, and land use changes.
IBM has also partnered with Environment and Climate Change Canada to test Prithvi's capabilities in precipitation nowcasting and high-resolution climate modeling. These collaborations highlight the model's versatility and its potential to drive advancements in climate science and public safety.
As climate change continues to present global challenges, open-source models like Prithvi aim to accelerate research and improve our understanding of complex weather and climate systems.
Prithvi and its specialized models are now accessible through Hugging Face, providing the global scientific community with powerful tools to enhance climate resilience and improve weather forecasting accuracy.