Researchers at the University of Manchester have retrained Nvidia's Earth-2 CorrDiff generative model to forecast air pollution across the United Kingdom at 2-3 square kilometer resolution, completing the training run in two days on a single eight-GPU node of the Isambard-AI supercomputer in Bristol. The project, led by physicist David Topping and doctoral student Hao Zhang, targets a public health problem tied to an estimated 30,000 UK deaths last year. The same workflow now runs inference on a desktop-class Nvidia DGX Spark system sitting in Topping's office.
Traditional chemistry-based air quality models are compute-heavy, which caps how finely they resolve pollution and how often they can be rerun. Topping's team sidestepped the chemistry bottleneck by generating training data from existing chemistry-climate simulations and feeding a year of UK pollution data at hourly intervals into Earth-2 CorrDiff, a generative downscaling model Nvidia originally built for weather. The model worked on the first attempt.
Topping framed the shift bluntly, arguing that pollution fields have enough in common with weather fields that Nvidia's generative frameworks translate directly.
Key facts
- 01University of Manchester trained Nvidia Earth-2 CorrDiff on a year of UK pollution data in two days on a single eight-GPU node of Isambard-AI.
- 02The resulting model forecasts UK-wide air pollution at 2-3 square kilometer resolution at hourly intervals.
- 03Air pollution contributed to an estimated 30,000 deaths in the UK last year, the health backdrop for the project.
- 04Isambard-AI is powered by 5,448 Nvidia GH200 Grace Hopper Superchips delivering 21 exaflops of AI performance.
- 05The same workflow now runs on an Nvidia DGX Spark desktop system, which Topping says costs a few thousand dollars to get started.
The training run used a fraction of Isambard-AI's total capacity. The UK's most powerful AI supercomputer packs 5,448 Nvidia GH200 Grace Hopper Superchips and delivers 21 exaflops of AI performance, but the pollution workload consumed only a single eight-GPU node for two days. Simon McIntosh-Smith, director of the Bristol Centre for Supercomputing at the University of Bristol and a co-founder of Isambard-AI, said the GPU hours used were relatively low, a notable outcome for a climate-adjacent project running on top-tier hardware.
The team has since layered on Earth-2 StormCast, a second Nvidia framework that produces time-dependent forecasts and can ingest live air quality observations. Zhang, who trained StormCast on Isambard-AI, said switching between the two Nvidia frameworks was straightforward, and that the group is only starting to explore how to combine them for more complex pollution fields.
The compute economics are the piece Nvidia is leaning into hardest. Once trained, the pollution model runs on the Nvidia GB10 Grace Blackwell-powered DGX Spark, a personal AI system Topping now uses for retraining smaller models and for inference. That collapses the barrier between national-scale supercomputing and desktop work.
Applications the group is planning include proactive alerts for healthcare systems — flagging asthma patients when pollution is forecast to spike in their area — and pairing the model with edge AI sensors for real-time response to events like wildfires. Topping also wants to run policy scenarios, modeling what happens to UK pollution under different regulatory changes.
Cost is the other headline. Topping said a researcher no longer needs institutional supercomputer access to get started, which reshapes who can build a national or municipal pollution model.
“You can now invest a few thousand dollars to get started developing powerful AI models.”— David Topping, Professor, University of Manchester
The team plans to open-source the training data and workflows so that other countries and cities can train equivalent models on their own local data with a short burst of supercomputer time. Topping's five-year horizon is an agentic interface: a clinician or government agency asking a natural-language question about tomorrow's neighborhood-level pollution and getting an answer from a chain of models grounded in the underlying physics.
The unresolved variable is data access. The forecast quality depends on open, high-quality air quality observations, which vary widely by country and by city. A workflow that runs on a DGX Spark is only as good as the sensor network feeding it, and many regions with the worst pollution have the thinnest monitoring coverage.
For Nvidia, the Manchester project is a proof point that the Earth-2 stack extends beyond weather into adjacent environmental domains without a rebuild, and that a national-supercomputer training run followed by desktop inference is a viable deployment pattern. If the open-source release lands as planned, the more interesting metric will not be the resolution of the UK model but the number of countries that stand up their own version within a year — a test of whether generative environmental modeling has actually decentralized, or whether the compute floor has simply moved from the cloud to a $4,000 box that still sits mostly in wealthy labs.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
