University of Manchester researchers have adapted Nvidia’s Earth-2 weather tools to model air pollution across the United Kingdom, replacing some of the expensive chemistry computation that limits traditional forecasts. The team trained the generative downscaling model CorrDiff on the Isambard-AI supercomputer and added StormCast for time-dependent predictions using observations.
Training used one year of hourly simulated UK pollution data and a single eight-GPU node. It took two days and produced a national model with cells roughly two to three kilometers wide. The inference workflow and smaller retraining runs can operate on Nvidia’s desk-side DGX Spark system. Researchers want to combine the model with live sensors, increase resolution toward street level and test policy scenarios such as changes to pollution controls. A possible health use would warn people with asthma before high-pollution periods.
The team plans to release training data and workflows so other regions can build models from local observations and a short allocation of supercomputer time. The results are presented by Nvidia and do not yet include an independent comparison with established chemistry-based forecasting systems. Speed and accessibility matter only if predictions remain accurate across weather, geography and unusual events such as wildfires. Open evaluation against measured pollution will therefore be the essential next step before health agencies rely on the output.