A UK University Trained an AI Air Pollution Forecaster in Two Days, Then Moved It to a Desktop

Key points

• University of Manchester researchers trained a UK-wide air pollution AI model in two days on Isambard-AI, the UK's national AI supercomputer, using a single eight-GPU node.
• The model covers the whole of the UK at 2 to 3 square kilometres resolution, using a full year of hourly pollution data.
• Training used Nvidia Earth-2 CorrDiff, an open AI model originally built for high-detail weather forecasting.
• The finished model runs on a DGX Spark desktop AI computer costing a few thousand dollars, so smaller research teams can use it without booking supercomputer time.
• The team plans to release training data and workflows as open source so other countries can build their own versions.

Air pollution kills an estimated 30,000 people in the UK every year. Scientists have long had the chemistry to model it, but the traditional approach is brutally slow: add chemical reactions to a weather simulation and computing costs spiral fast. David Topping, a professor of Earth and environmental science at the University of Manchester, decided to try a shortcut.

He looked at what Nvidia's Earth-2 project, a family of open AI tools built for high-resolution climate and weather forecasting, had already done for predicting storms. The core idea was to use a generative AI model, a type of AI that learns patterns from data and can produce detailed outputs from rough inputs, to fill in the fine-grained pollution picture that expensive chemistry models normally provide.

It worked on the first attempt.

How did they actually build it?

The team trained on Isambard-AI, housed in Bristol, which packs 5,448 Nvidia GH200 Grace Hopper Superchips and delivers 21 exaflops of AI computing power (an exaflop is roughly a billion billion calculations per second). One eight-GPU node and a year's worth of simulated UK pollution data at hourly intervals was all it took. Two days later, they had a model.

They used two Nvidia Earth-2 tools. CorrDiff is a downscaling model: it takes a low-resolution pollution map and sharpens it to street-level detail. StormCast handles time, letting the model ingest real air quality readings and project them forward into forecasts.

The same workflow transferred to a DGX Spark, a desktop machine powered by Nvidia's GB10 Grace Blackwell superchip. Topping now has one in his office and uses it to retrain models. "You can now invest a few thousand dollars to get started developing powerful AI models," he told Nvidia. Our story on 3 September found Nvidia already pushing spare compute onto everyday hardware with its PAIR software; this project shows the same logic applied to serious research infrastructure.

What does this mean for ordinary people?

The practical ambition is direct. Topping wants regional health services to warn asthma patients a day or a week ahead when pollution in their area is forecast to spike. The model can also simulate policy scenarios, such as what happens to air quality if a city bans diesel vehicles or if a new industrial site opens.

Longer term, the team's exploring pairing the model with edge AI devices, small sensors that process data locally rather than sending everything to a remote server, to handle fast-moving events like wildfires in near real time.

Training data and workflows will be released as open source. Topping's stated goal, as reported by Nvidia, is that "every global country and every major city" could build its own pollution model with a short burst of supercomputer time and local data. Given what ex-EPA officials warned, in a piece Nvidia reported on 17 September, about deregulation and public health, tools that let smaller institutions run their own air quality science matter more than they might look at first glance. AI2Day covered those warnings in "Former EPA Officials Say Deregulation Is Turning the AI Boom Into a Public Health Crisis".

The honest caveat: it's still a research prototype. A clinician can't query it today. But the gap between a two-day training run and a public health tool is narrowing faster than most people expected.

Common questions

Does the model work outside the UK?

Not yet as a ready-made product, but the open-source release is designed exactly for that. Any country with local pollution data and a short run on a national AI supercomputer should be able to train its own version using the same workflow.

How accurate is 2 to 3 square kilometres?

Detailed enough to distinguish between a busy main road and a quieter street a few blocks away, which is far more granular than most current public air quality forecasts. The team plans to push resolution further by adding more open data sources.

Can a hospital or GP surgery use this right now?

Not directly. The current work is at the research stage. Topping's vision of an AI a clinician can simply ask about tomorrow's pollution in a specific neighbourhood is described as a five-year horizon.