Rapid Training on Isambard-AI Infrastructure
Researchers at the University of Manchester successfully trained a new air pollution forecasting model using NVIDIA Earth-2 technology. The initial training process was completed in just two days, utilizing a single eight-GPU node on Isambard-AI, the U.K.’s most powerful AI supercomputer. This infrastructure is equipped with 5,448 NVIDIA GH200 Grace Hopper Superchips that deliver 21 exaflops of AI performance. The team noted that the model worked effectively on its first attempt.
Photo: Coolcaesar · CC BY-SA 4.0
High-Resolution Spatial Data Generation
To create a detailed U.K.-wide pollution model, the team used a year’s worth of simulated pollution data recorded at hourly intervals. This dataset allowed them to generate a nationwide model with a spatial resolution of roughly two to three square kilometers. By leveraging existing climate simulations, the researchers effectively taught the AI to predict pollution patterns without relying on slow chemical calculations. The resulting model provides a comprehensive look back at air pollution over the past year.
Integration of Generative AI Frameworks
The project applies generative AI techniques originally developed for weather and climate modeling to air quality forecasting. Physicists David Topping and Hao Zhang trained NVIDIA's existing Earth-2 CorrDiff model, a generative downscaling framework, using data from the U.K.'s Isambard-AI supercomputer. The team also incorporated Earth-2 StormCast, which enables time-dependent forecasts by directly using air quality observations.
Addressing Significant Public Health Risks
The development of this forecasting model addresses a critical public health issue in the United Kingdom. Air pollution is estimated to contribute to approximately 30,000 deaths in the U.K. alone each year. By enhancing the accuracy and frequency of air quality predictions, the integration of NVIDIA’s generative frameworks aims to provide better data for public health interventions. The model can help predict future air pollution scenarios, offering valuable insights for policymakers.
Future Scalability and Localized Applications
The team plans to further increase the model's resolution by incorporating additional open data, which would enable researchers to understand air pollution at a street scale. NVIDIA indicates that the same workflow can now perform inference and smaller training runs on a DGX Spark desktop system powered by the GB10 Grace Blackwell superchip. The longer-term goal is to allow other countries and cities to create similar models using their own local pollution observations and relatively short bursts of supercomputer training time.
Predictive Capabilities Beyond Historical Data
While the initial model provides a detailed look back at air pollution over the past year, it is also designed to predict future air pollution scenarios for the United Kingdom. This forward-looking capability allows researchers to move beyond simple historical reconstruction and actively forecast upcoming environmental conditions. By combining retrospective analysis with predictive modeling, the system offers a more comprehensive tool for understanding how pollution levels might evolve in specific regions.
Eliminating Slow Chemical Calculations
The team generated training data from existing climate simulations, effectively teaching the AI to predict pollution patterns without relying on exhaustive and slow chemical calculations. This approach bypasses traditional computational bottlenecks associated with complex atmospheric chemistry modeling. By leveraging generative frameworks originally developed for weather and climate fields, the researchers created a more efficient pathway to high-resolution air quality forecasting.
- University of Manchester Uses NVIDIA Earth-2 to Forecast Air Pollution Across the UK | NVIDIA Blog
- Children’s Hospital Uses Open Source NVIDIA AI for Cardiac Care | NVIDIA Blog
- NVIDIA Earth-2 Speeds Up UK-Wide Air Pollution Forecasting – Techtroduce
- Manchester Physicists Use AI To Predict UK Air Quality
- University of Manchester Leverages NVIDIA Earth-2 for UK Air Pollution Forecasting – RobotToday