University of Manchester Uses NVIDIA Earth-2 for UK Pollution Forecasts
University of Manchester adapts NVIDIA Earth-2 to forecast UK air pollution at 2 to 3 square kilometer resolution after a two-day run on Isambard-AI.
Summary
Published September 16, 2026, the University of Manchester project targets air pollution, which contributed to an estimated 30,000 UK deaths last year, without the cost and slowness of chemistry-based forecasting. Earth and environmental science professor David Topping worked with NVIDIA’s Earth-2 team to generate training data from chemistry-climate simulations and adapt the generative CorrDiff downscaler. One year of hourly simulated UK pollution data produced a nationwide model at 2 to 3 square kilometer resolution. CorrDiff worked on its first attempt and trained in two days on one eight-GPU Isambard-AI node. The University of Bristol supercomputer contains 5,448 NVIDIA GH200 Grace Hopper Superchips delivering 21 exaflops. Bristol Centre for Supercomputing director and Isambard-AI cofounder Simon McIntosh-Smith said the run used relatively few GPU hours and less power.
University of Manchester doctoral student Hao Zhang then trained Earth-2 StormCast, adding time-dependent forecasts that directly use observations. Test-training and inference workflows ran on DGX Spark, whose NVIDIA GB10 Grace Blackwell superchip supports inference and smaller retraining runs from Topping’s office for an investment of a few thousand dollars. NVIDIA developer relations manager Niall Robinson said moving the workflow from a national supercomputer to a desktop broadens access. Proposed uses include testing government policies, warning asthma patients about pollution tomorrow or next week, and combining edge AI data with the model for real-time wildfire decisions. The team plans to add open data for street-scale resolution and release its training data and workflows so countries and cities can build local models. Within five years, Topping envisions an agentic interface that lets clinicians or agencies query a chain of science-grounded models.
Positives
- CorrDiff trained in two days on one eight-GPU Isambard-AI node and worked on its first attempt.
- One year of hourly data produced nationwide pollution modeling at 2 to 3 square kilometer resolution.
- DGX Spark can run inference and smaller retraining jobs from a desktop system costing a few thousand dollars.
- StormCast adds time-dependent forecasts that directly incorporate air quality observations.
- Open source training data and workflows could let countries and cities create models using local pollution data.
- Policy testing, asthma alerts and real-time wildfire decisions could turn forecasts into public health action.
Risks & concerns
- Air pollution contributed to an estimated 30,000 UK deaths last year.
- Chemistry-based air quality models remain expensive and slow, limiting their detail and update frequency.
- Street-scale resolution still requires additional open data and further model development.
- Real-time decisions using edge AI air quality devices remain under exploration rather than deployment.
- Global replication requires local data and at least a short allocation of supercomputer AI capacity.