A University of Manchester team has adapted NVIDIA’s Earth-2 generative AI tools to forecast air pollution across the UK at a resolution of 2–3 square kilometres. The practical promise is sharper local warnings and faster testing of pollution policy, although the results and future benefits described here remain claims from NVIDIA and the research team rather than an independent assessment.
NVIDIA Watch analysis
What happened
Professor David Topping and colleagues trained NVIDIA’s Earth-2 CorrDiff model using a year of hourly UK pollution data generated from existing chemistry-climate simulations. NVIDIA says the training took two days on a single eight-GPU node at the Isambard-AI supercomputer in Bristol.
The team has also added Earth-2 StormCast, which can use air-quality observations for time-dependent forecasts. NVIDIA says the workflow runs on its DGX Spark desktop system for inference and smaller training jobs, putting a version of the work on hardware that is considerably less exotic than a national supercomputer. Still not pocket money, obviously, but a meaningful step down from “only a major facility can attempt this”.
Why it matters
Air-quality modelling is computationally expensive because it combines weather, chemistry and geography. Faster modelling could help researchers test what might happen under different pollution policies, while health services could eventually use local forecasts to warn people with conditions such as asthma.
The model is not yet a finished public warning service. The team plans to improve the resolution towards street scale, add more open data and release training data and workflows so researchers in other countries can adapt the approach. That makes the important development less about one NVIDIA demonstration and more about whether environmental forecasting becomes easier to reproduce locally.
Our read
This is a solid example of AI doing useful, slightly unglamorous work: not writing a sonnet, but helping estimate what people may breathe tomorrow. The strongest supported point is the combination of speed, geographical detail and portability. The biggest unanswered question is whether the model remains reliable when it leaves the training conditions and meets messy real-world observations.
For now, treat the claimed performance and health applications as a promising research direction, not a validated public-health prediction system.
What to watch
- Independent validation
Whether the pollution forecasts are tested against measured observations across different seasons and regions. - Street-scale accuracy
Whether adding open data genuinely improves neighbourhood-level predictions rather than merely producing finer-looking maps. - Open release
Whether the promised training data and workflows are published, documented and usable beyond NVIDIA hardware. - Public deployment
Whether health services or government agencies turn the research into operational warnings people can act on.
Discussion spark: Should public-health forecasting systems prioritise a fast, reproducible AI model, or wait for slower chemistry-based methods to establish stronger independent validation?
Sources and evidence
- University of Manchester Uses NVIDIA Earth-2 to Forecast Air Pollution Across the UK – NVIDIA Blog (16 September 2026, 05:06 UTC)
Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.