University of Manchester runs national air pollution forecast on NVIDIA Earth-2

Air pollution causes roughly 30 thousand deaths a year in the UK alone, yet chemistry-based models that produce air-quality forecasts remain too expensive to run at high resolution and useful frequency. Professor David Topping of the University of Manchester's Department of Earth and Environmental Sciences recognised that NVIDIA's open Earth-2 model family had already cracked a similar problem in weather forecasting, and asked whether the same generative frameworks could work for pollution fields. Once chemistry is added to weather models, he said, they become "really, really slow", making the generative route a natural choice.

The team generated training data from existing chemistry-climate simulations, then trained Earth-2 CorrDiff, a generative downscaling model, on Isambard-AI, the UK's national AI supercomputer in Bristol. The model succeeded on the first attempt. Training drew on one year of UK pollution data at hourly intervals and produced a national model at 2-3 km² resolution. On a single eight-GPU node of Isambard-AI — which houses 5,448 GH200 Grace Hopper Superchips and delivers 21 exaflops (21 billion billion floating-point operations per second) of AI performance — the process took just two days. Simon McIntosh-Smith, director of the Bristol Centre for Supercomputing and a co-founder of Isambard-AI, noted that CorrDiff showed exceptional hardware efficiency, with GPU hours required relatively low for a climate project.

Since then the team has added Earth-2 StormCast, a model that enables time-dependent forecasts incorporating actual air-quality observations, and demonstrated both training and inference workflows on NVIDIA DGX Spark, a personal AI compute workstation. Niall Robinson, NVIDIA's head of developer relations for weather and climate, said the shift from a two-day training run on a national supercomputer to inference on a desk-side device changes who can do this science and how quickly. PhD student Hao Zhang, who trained StormCast on Isambard-AI, said the flexibility of moving between NVIDIA's different frameworks impressed him, and that the team is only beginning to explore further applications for complex pollution fields.

Topping describes a scenario in which regional and national health services could alert asthma patients in advance of expected pollution tomorrow or next week, and use the model to test policy scenarios — for example, what would happen if different regulatory changes were implemented. Another direction under investigation is integration with edge AI devices to ingest real-time air-quality data and support immediate decisions, such as during wildfires. Robinson concludes that this is only the beginning of what these open workflows can enable on a global scale.