University of Manchester researchers have adapted NVIDIA Earth-2 generative AI frameworks to forecast air pollution across the United Kingdom, NVIDIA announced. The project targets an environmental hazard that contributed to an estimated 30,000 deaths in the U.K. last year.
Traditional chemistry-based models require massive computing power, which restricts how often scientists can run them and limits their spatial detail. David Topping, a professor in the university’s department of Earth and environmental science, collaborated with the NVIDIA Earth-2 team to test whether generative climate and weather models could simulate pollution fields instead.
The team generated training data from existing chemistry-climate simulations, covering a year of U.K. pollution figures simulated at hourly intervals. They then trained Earth-2 CorrDiff, a generative downscaling model, to generate a nationwide pollution model at a resolution of 2 to 3 square kilometers. Training ran on a single eight-GPU node on Isambard-AI, the national AI supercomputer located in Bristol. Isambard-AI houses 5,448 NVIDIA GH200 Grace Hopper Superchips delivering 21 exaflops of AI performance. The training run took two days and succeeded on its first attempt.
Supercomputing to the Desktop
Researchers expanded the project by adding Earth-2 StormCast, an AI model that produces time-dependent forecasts directly from air quality observations. Hao Zhang, a doctoral student at the University of Manchester, trained StormCast on Isambard-AI. The researchers also ran test-training and inference workflows on the NVIDIA DGX Spark, a desktop AI system powered by the GB10 Grace Blackwell superchip.
The team intends to use the models to evaluate potential future scenarios, including the impact of government policy changes. Healthcare organizations could also use the forecasts to issue early warnings to patients with respiratory conditions such as asthma before pollution spikes occur. Researchers are also exploring integrations with edge AI devices to ingest real-time air quality data during sudden events like wildfires.
Topping and his team plan to release the training data and workflows as open source so other countries and municipalities can build localized models. The researchers also plan to incorporate additional open data to scale down model resolution to individual streets, with a longer-term five-year goal of building an agentic query interface for clinicians and public agencies.
