Manchester researchers forecast UK air pollution with NVIDIA Earth-2
A University of Manchester team retrained Earth-2 models on pollution data, completing training in two days on the Isambard-AI supercomputer.
A team at the University of Manchester has adapted NVIDIA's open Earth-2 weather and climate models to forecast air pollution across the UK. David Topping, a professor in the department of Earth and environmental science, points to cost as the core obstacle: "Once you put chemistry into weather models, they get really, really slow."
The researchers 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. Using a year of UK pollution data simulated at hourly intervals, they produced a UK-wide model at a resolution of 2-3 square kilometres.
Two days of training
The run took two days on a single eight-GPU node of Isambard-AI, which packs 5,448 NVIDIA GH200 Grace Hopper Superchips and delivers 21 exaflops of AI performance. The same workflow also runs on the GB10 Grace Blackwell-powered DGX Spark desktop system for inference and smaller training runs, and Topping now retrains models on one in his office. "You can now invest a few thousand dollars to get started developing powerful AI models," he said.
The team later added Earth-2 StormCast, which enables time-dependent forecasts that draw directly on air quality observations. Niall Robinson, developer relations manager for weather and climate at NVIDIA, said the short training time and desktop inference "changes who can do this science and how quickly."
Where it could be used
Topping describes healthcare services alerting patients with conditions such as asthma when pollution is expected to spike locally. The model can also simulate the effect of different government policy scenarios. The team is exploring real-time data from edge AI devices, for cases such as wildfires, and plans to raise resolution to street scale using additional open data.
Training data and workflows are to be released as open source so other countries can build equivalent models. Full details are on the NVIDIA blog.
