IBM and NASA launch innovative open-source AI model for weather forecasting

IBM and NASA have jointly unveiled a groundbreaking open-source artificial intelligence (AI) model designed to revolutionise weather and climate forecasting. This collaboration, which also included contributions from Oak Ridge National Laboratory, introduces a flexible and scalable AI foundation model poised to tackle diverse challenges in short-term weather prediction and long-term climate projection. This development is significant not only for meteorologists but also for developers, scientists, and business stakeholders engaged in climate-related fields.

The model, termed “Prithvi WxC,” stands out due to its innovative architecture and training methodology, allowing it to address a more comprehensive array of applications compared to existing weather AI models. According to a detailed report in a paper on arXiv, potential uses span targeted forecasting using local observations, identifying and predicting extreme weather patterns, enhancing the spatial resolution of global climate models, and refining the depiction of physical processes within numerical weather and climate simulations.

Notably, during an experimental phase, this AI model demonstrated its capability by accurately reconstructing global surface temperatures using just five percent of the original data, highlighting its potential in data assimilation challenges. The model’s foundation derives from pre-training on 40 years of Earth observation data, sourced from NASA’s Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2). This extensive dataset underpins its ability to adapt to global, regional, and local scales—making it a versatile tool for a plethora of weather-related studies.

Two specific fine-tuned versions of the model are now available alongside the foundation model on Hugging Face, a platform for open AI models. One of these is dedicated to climate and weather data downscaling—a crucial meteorological practice that enables high-resolution forecasts from low-resolution input data like temperature, precipitation, and wind conditions. This fine-tuned model can project weather and climate data with up to twelve times greater resolution, offering detailed forecasts and climate projections.

The second specialised model focuses on gravity wave parameterization. Gravity waves, prevalent in the atmosphere, influence several atmospheric occurrences such as cloud formation and aircraft turbulence. Traditionally, numerical models have struggled to accurately capture gravity waves, leaving a degree of uncertainty regarding their climate impacts. This model aims to enhance scientific understanding by improving estimates of gravity wave generation, precision in numerical weather models, and reducing uncertainty in future climate event simulations.

The initiative has not gone unnoticed by influential figures. Karen St. Germain, Director of NASA’s Earth Science Division within its Science Mission Directorate, emphasised the significance of the model in meeting the pressing demands of contemporary climate challenges, stating that it provides actionable science useful to individuals and organisations in preparing and responding to climatic events.

Juan Bernabe-Moreno, Director of IBM Research Europe (UK and Ireland) and IBM’s Accelerated Discovery Lead for Climate and Sustainability, explained that the adaptability of the model means it can be applied both globally and locally, enabling better understanding and predictions of meteorological phenomena, including hurricanes and atmospheric rivers.

The launch of this model is part of a broader collaborative effort by IBM Research and NASA to harness AI technology in understanding Earth’s climatic systems. The partners aim to enable crucial research breakthroughs, particularly in computing and data, to address significant national weather and climate issues.

Preliminary applications have already commenced, with Environment and Climate Change Canada (ECCC) testing the model’s capabilities. They are investigating short-term precipitation forecasts using a technique called precipitation nowcasting, which integrates real-time radar data, and examining downscaling from global forecasts to km-scale resolution.

This latest weather and climate model extends the Prithvi family of AI foundation models, a testament to the ongoing partnership between IBM and NASA. The foundation model, alongside the gravity wave parameterization model, can be accessed via the NASA-IBM page on Hugging Face, while the downscaling model is available on the IBM Granite page on the same platform. This initiative follows a similar release last year of the largest open-source geospatial AI model, which has significantly supported governments, companies, and institutions in analysing changes in disaster patterns and other geophysical processes.

Source: Noah Wire Services