Summary
University of Manchester researchers used NVIDIA Earth-2 generative models to build a UK-wide air-pollution forecasting workflow at 2–3 kilometre resolution. The training process took two days on an eight-GPU node at Isambard-AI, while smaller training runs and inference now run on NVIDIA’s DGX Spark desktop system.
University of Manchester researchers have adapted NVIDIA’s Earth-2 generative AI models to forecast air pollution across the UK, according to an NVIDIA account of the project. The resulting workflow produces pollution fields at a resolution of 2–3 square kilometres and was trained in two days on an eight-GPU node at the Isambard-AI supercomputer in Bristol.
The work is intended to make detailed air-quality modelling faster and easier to adapt. The team is also developing time-dependent forecasts that use air-quality observations and is exploring applications ranging from policy analysis to advance warnings for healthcare services.
Contents
- How the pollution model was built
- From national supercomputer to desktop system
- Potential uses and current scope
How the pollution model was built
Air-quality forecasts require more than weather information. Chemistry-based models must represent how pollutants are emitted, transported and transformed in the atmosphere. The NVIDIA post says these calculations can make conventional air-quality forecasting computationally expensive.
David Topping, a professor in the University of Manchester’s Department of Earth and Environmental Science, proposed applying the generative frameworks used in Earth-2 to pollution fields. His team generated training data from existing chemistry-climate simulations, using one year of UK pollution data simulated at hourly intervals.
The researchers first trained Earth-2 CorrDiff, which NVIDIA describes as a generative downscaling model. Downscaling uses a model’s broader atmospheric information to produce a more detailed field at a finer spatial resolution. The resulting pollution model covers the UK at 2–3 km resolution.
The team later added Earth-2 StormCast for time-dependent forecasts that can directly use air-quality observations. The source says the training and inference workflows were run on Isambard-AI, the UK national AI supercomputer.
From national supercomputer to desktop system
The initial training process ran on a single eight-GPU node and took two days. NVIDIA describes Isambard-AI as containing 5,448 GH200 Grace Hopper superchips and providing 21 exaflops of AI performance.
The same pollution-modelling workflow also runs on the DGX Spark desktop AI system, which is powered by NVIDIA’s GB10 Grace Blackwell superchip. The system is being used for inference and smaller training runs, and Topping is retraining models on a DGX Spark in his office.
That shift changes the computing requirements for experimentation. A national supercomputer is still used for the major training workload, but researchers can use a much smaller local system to test models, run forecasts and carry out limited retraining.
Potential uses and current scope
The model can represent pollution over the past year and support future scenario modelling. One proposed application is testing how pollution patterns could change if government policies were modified. This makes the workflow potentially useful for examining policy scenarios before they are implemented.
The team also envisions proactive alerts for healthcare organisations. Regional or national services could eventually use forecasts to notify people with conditions such as asthma when pollution is expected to be high in their area. The project is further exploring the use of edge-AI devices to supply real-time air-quality measurements, including during events such as wildfires.
The source describes these healthcare and real-time-response applications as potential or ongoing areas of work. The current model is reported at 2–3 km resolution, while the team plans to move towards street-scale modelling by incorporating additional open data. The researchers also plan to release open-source training data and workflows so that similar pollution models can be developed for other countries and cities using local data.
This is therefore a model-development result rather than a reported public-health deployment. The announcement gives training speed, computing requirements and intended applications, while a comparison of forecast accuracy with conventional chemistry-based systems is not provided. The practical value of the system will depend on the quality of local observations, the performance of the forecasts and how public agencies integrate them into decisions.