IBM and NASA unveil open source AI climate model promising efficient weather predictions

IBM and NASA Unveil Open Source AI Climate Model Promising Efficient Weather Predictions

This week, researchers from IBM and NASA have debuted a revolutionary AI climate model named Prithvi WxC, promising to significantly enhance the efficiency and accuracy of weather prediction. This innovative development comes as the result of an extensive collaboration, which also involved the US Department of Energy’s Oak Ridge National Laboratory.

Prithvi WxC, a foundation model embedding 2.3 billion parameters, has been trained on four decades of observational data drawn from NASA’s MERRA-2 dataset. Despite its compact scale, the model has demonstrated an impressive capability to generate global surface temperature predictions using a mere 5% sample of the original data set. Researchers assert that Prithvi WxC holds particular proficiency in simulating complex weather phenomena such as hurricanes and atmospheric rivers.

A distinguishing feature of Prithvi WxC lies in its versatility. Unlike other AI models that are restricted to a single dataset and use case, Prithvi WxC is designed to be adaptable for various applications. This flexibility allows it to be tailored for both short-term weather forecasting and long-term climate projections—a notable advancement in the realm of AI in climatology.

According to Juan Bernabe-Moreno, Director of IBM Research Europe, “We have designed our weather and climate foundation model to go beyond such limitations so that it can be tuned to a variety of inputs and uses.”

To facilitate broader adoption and further innovation, IBM and NASA have made Prithvi WxC available on Hugging Face. This release includes two fine-tuned models focused on specific tasks: climate and weather downscaling, and gravity wave parameterization. Weather downscaling enhances low-resolution data to produce high-resolution forecasts, while gravity wave parameterization addresses atmospheric processes impacting cloud formation and aircraft turbulence.

Karen St Germain, Director of NASA’s Earth Science Division, emphasised the model’s potential utility, stating that it “will help us produce a tool that people can use [for] weather, season and climate projections to help inform decisions on how to prepare, respond, and mitigate.”

Prithvi WxC’s relatively modest computational requirements are another key advantage. The model was trained using just 64 Nvidia A100 GPUs, and future fine-tuning is expected to demand even fewer resources. This accessibility could democratize advanced weather modeling, allowing climate centres worldwide to integrate and adapt the model using their own supercomputing resources.

One of the early adopters of Prithvi WxC is the Canadian government. Environment and Climate Change Canada (ECCC) is working to incorporate the model for enhanced weather forecasting. Specifically, ECCC aims to improve short-term precipitation forecasts by integrating real-time radar data and refine predictions through downscaling techniques to achieve forecasts accurate to the kilometre.

Prithvi WxC stands as a milestone in AI-driven climatology, promising to bridge gaps between data complexity, computational efficiency, and forecasting accuracy. As other research entities like Google and Nvidia continue to explore AI for climatology, the release of Prithvi WxC could herald a new era of sophisticated, resource-efficient climate modeling.

Source: Noah Wire Services