Manchester team adapts NVIDIA Earth-2 for UK pollution forecasts

Air pollution contributed to an estimated 30,000 deaths in the U.K. last year, and University of Manchester researchers are using NVIDIA Earth-2 tools to make forecasting less compute-intensive. Led by David Topping, the team trained Earth-2 CorrDiff on chemistry-climate simulation data to generate U.K.-wide pollution fields, adapting generative weather and climate frameworks for air quality modeling.

The work ran on Isambard-AI, the U.K.’s national AI supercomputer in Bristol, where a year’s worth of hourly simulated U.K. pollution data produced a model at a resolution of 2-3 square kilometers. Training on a single, eight-GPU node took two days on a system with 5,448 NVIDIA GH200 Grace Hopper Superchips and 21 exaflops of AI performance.

The team has added Earth-2 StormCast for time-dependent forecasts that use air quality observations directly, and has shown workflows running on NVIDIA DGX Spark for inference and smaller training runs. Potential uses include health alerts for asthma patients, real-time wildfire response using edge AI data, policy scenario modeling, and eventual street-scale pollution analysis through additional open data.

Researchers plan to release open source training data and workflows so other countries and cities can build models with local data. Topping’s longer-term goal is an agentic interface that lets clinicians or government agencies ask neighborhood-level pollution questions and receive science-grounded forecasts from a chain of models.