University of Manchester Develops UK-Wide Air Pollution Prediction Model Using NVIDIA Earth-2

A team led by Professor David Topping from the Department of Earth and Environmental Science at the University of Manchester developed an air pollution prediction model for the entire UK using the NVIDIA Earth-2 open AI models and tools. While existing chemistry-based models had limitations in detail and execution frequency due to high computational costs, this approach improved efficiency by applying a generative framework previously used for weather forecasting to pollution fields.

Technical Implementation and Performance

The research team trained the Earth-2 CorrDiff model on Isambard-AI, the UK's national AI supercomputer located in Bristol. The model was trained on one year of UK pollution data to perform high-resolution predictions at 2–3 square kilometer scales, completing training in just two days on a single 8-GPU node. Additionally, the time-dependent prediction model Earth-2 StormCast, which directly uses air quality observation data, was added to complement prediction accuracy.

Deployment and Scalability

The developed model supports inference and small-scale retraining on the NVIDIA DGX Spark (a desktop AI system based on the GB10 Grace Blackwell superchip). Professor Topping emphasized that powerful AI models can be developed with an investment of thousands of dollars, predicting that the shift from supercomputers to desktops will change the actors and speed of research.

Future Plans and Applications

The research team plans to integrate additional open data to improve resolution to the street scale. The model is also expected to be used for policy scenario analysis, localized air pollution pre-alerts for asthma patients, and emergency response to situations such as wildfires. An Open Science project is also underway to release open-source training data and workflows so that similar models can be trained in other countries.