IBM and NASA Launch AI Model for Weather and Climate Predictions

The prospects for AI-enhanced weather forecasting are bright. Certain deep learning models, which have been trained on historical weather records, can already rival conventional weather models reliant on large-scale supercomputers that are often inaccessible to the average individual.

In partnership with NASA and contributions from Oak Ridge National Laboratory, IBM has created Prithvi WxC, a customizable, open-source foundational model for applications related to weather and climate. As stated by IBM Research, the model is optimized to run on desktop computers and is available on Hugging Face.

The training process for the model involved several weeks and numerous GPUs to analyze 40 years of historical weather data sourced from NASA’s MERRA-2 harmonized dataset, comprising satellite and various Earth observation data. The model is designed for swift tuning across different scenarios and can be deployed from a desktop in mere seconds. Potential uses encompass generating localized forecasts from weather data, anticipating severe weather events, augmenting the spatial accuracy of global climate simulations, and refining the depiction of physical phenomena in traditional weather and climate models.

“We developed our foundational model with the intention that the significant effort and GPU time invested initially would enable users to quickly derive and execute new applications,” explained Campbell Watson, an IBM climate researcher involved in the model’s development.

In a notable experiment, the model accurately reconstructed global surface temperatures using a minuscule, localized sample of weather data, filling in 95% of missing values. “The capability to generalize from a small, high-quality historical data sample to the global scale proves advantageous for various weather and climate projection tasks,” mentioned Juan Bernabé-Moreno, the director of IBM Research Europe and lead for climate and sustainability.

Localized Forecasting, Hurricane Tracking, and Understanding Earth’s Gravity Waves

The newly developed weather and climate foundational model is elaborated in a recent paper available on arXiv. Researchers outlined the methodology used to develop the model and fine-tune it with specialized data to create three applications of immediate relevance for weather forecasters.

The initial application focuses on enhancing low-resolution data to provide greater detail, a technique known as downscaling. By refining weather and climate predictions, downscaling can offer warnings for severe flooding events or intense hurricane winds. IBM’s downscaling application elevates data across diverse resolutions and types, such as temperature and precipitation, magnifying them by as much as 12 times. This downscaling tool is accessible via IBM’s Granite geospatial models on Hugging Face.

The second application is centered on hurricane forecasting. Researchers utilized the model to accurately track the path of Hurricane Ida, which impacted Louisiana in 2021, incurring damages of $75 billion, making it the fourth most expensive Atlantic hurricane recorded. Moving forward, this model could enhance the precision of identifying areas where defenses against hurricanes should be reinforced.

The third application aims to improve assessments of gravity waves. In the atmosphere, gravity waves affect cloud formation and global weather patterns, influencing points of aircraft turbulence. Traditional climate models tend to inadequately capture high-resolution gravity waves, leading to uncertainties in weather and climate forecasts. This could notably transform the management of global supply chains.

Additionally, IBM is collaborating with Environment and Climate Change Canada to tailor the foundational model for precipitation nowcasting, which involves utilizing real-time radar data to provide precise predictions for local rainfall several hours in advance. The objective is to see if this data-driven foundational model approach can use fewer computational resources while delivering enhanced accuracy.

Evolving to Think Like Meteorologists

This new foundation model for weather and climate adds to a growing suite of open-source models that aim to streamline the analysis of NASA’s satellite and other Earth observation data. Its versatility stems from a hybrid architecture combined with a unique training method.

Built using a vision transformer and a masked autoencoder, the model can encode spatial data as it unfolds over time. By extending its attention mechanism to incorporate time, it processes MERRA-2 reanalysis data, which consolidates various streams of observational data.

Moreover, the model can operate on both a spherical dimension, akin to traditional gridded climate models, and on a flat, rectangular surface. This dual representation enables a seamless transition from global to regional perspectives without compromising detail.

During the training phase, researchers presented the model with heavily obscured climate reanalysis datasets, instructing it to reconstruct images pixel by pixel, alongside projecting these blacked-out images into the future. “The model effectively learns how the atmosphere changes over time,” noted Johannes Schmude, an IBM researcher collaborating on the model’s development.

By prompting the model to fill in incomplete weather data and conceptualize its future state, researchers achieved two key outcomes: halving the dataset required for training, thus reducing GPU and energy demand, and teaching the model to infer missing information in both present and future contexts. This process mimics the tasks performed by meteorologists.

“Weather data is intrinsically sparse,” remarked Schmude. “To forecast accurately, one must learn to fill in the gaps.”

Future Directions

IBM and NASA plan to explore the integration of their existing open-source geospatial AI model for analyzing earth observation data with their new weather and climate model. The Prithvi Earth Observation model, launched last year, has been transformed into a range of applications that have collectively achieved over 10,000 downloads. These applications have been utilized, among other tasks, to assess the extent of previous flooding and gauge the severity of past wildfires using burn scar analysis.

Together, the Earth Observation model and the weather and climate model could tackle similarly complex challenges, from projecting crop yields to anticipating severe flooding incidents and their effects on communities.

For more information, visit the original source at IBM Research.

Image source: Shutterstock