Manchester University Leverages NVIDIA Earth‑2 to Model UK Air Quality

David Topping and Hao Zhang, senior scientists at the University of Manchester, joined forces with the Earth‑2 team at NVIDIA to adapt two open‑source generative models, CorrDiff and StormCast, for the purpose of forecasting air quality across the United Kingdom. Their work builds directly on a comprehensive, year‑long dataset that captures British atmospheric chemistry and climate dynamics, with observations recorded each hour and mapped onto a grid whose cells span roughly two to three kilometres. By leveraging this fine‑grained input, the researchers are able to train models that can resolve local variations in pollutant concentrations, something that traditional national‑scale forecasts often miss.

The training phase for the CorrDiff model was deliberately kept short: it required only two days of compute on a single node of the Isambard‑AI supercomputer located in Bristol. That node is equipped with eight GPUs, forming part of a broader installation that aggregates 5 448 GH200 superchips and advertises a peak AI performance of 21 exaflops. This modest training window illustrates how a dedicated research cluster can accelerate prototype development without needing to commandeer the entire supercomputing resource, thereby preserving capacity for other scientific workloads.

From Prototype to Operational Service

At present, the CorrDiff model is capable of generating high‑resolution concentration fields for key pollutants such as nitrogen dioxide (NO₂) and fine particulate matter (PM₂.₅). The StormCast extension enriches this capability by adding a weather‑driven component that ingests real‑time air‑quality observations from monitoring stations. By fusing observational data with meteorological inputs, StormCast can produce short‑term forecasts that are immediately relevant to vulnerable populations, for instance people with asthma who depend on timely alerts to manage their exposure.

The research team is careful to stress that the current system remains a prototype. Inference and occasional fine‑tuning are performed on a DGX Spark platform that incorporates a Grace Blackwell (GB10) accelerator, rather than on a conventional desktop workstation. This distinction matters because it prevents stakeholders from assuming that a single high‑end PC could replace the need for a dedicated supercomputing environment, which provides the necessary throughput and memory bandwidth to handle the model’s computational demands.

Potential Applications

• Real‑time alerts for asthma patients and other sensitive groups
• Scenario testing for public‑policy interventions such as low‑emission zones
• Integration with peripheral sensor networks to fill gaps in monitoring coverage
• Support for emergency response during wildfire events by predicting smoke dispersion

Looking beyond these immediate uses, the team intends to release both the training dataset and the workflow scripts to the wider research community. Their longer‑term ambition is to push the spatial resolution down to the street level, a goal that will require the incorporation of additional open data sources and more granular emission inventories. By making the resources openly available, they hope to stimulate collaborative improvements and enable other groups to tailor the models to local contexts.

Why Better Forecasts Matter

According to the Royal College of Physicians, air‑pollution‑related mortality in the UK is projected to reach roughly 30 000 deaths in 2025, imposing an economic burden of over £27 billion each year. The College underscores that no level of exposure can be considered entirely safe and that the health impacts are disproportionately borne by disadvantaged socioeconomic groups. Improved forecasting therefore becomes a public‑health imperative, offering the possibility of early warnings that can mitigate exposure for those most at risk.

For organisations operating in English‑speaking markets, the emerging capability to generate fine‑grained air‑quality forecasts introduces a new layer of situational awareness. Energy utilities could, for example, adjust grid loads in anticipation of pollution‑driven demand spikes, while insurers might refine risk models for health‑related claims based on forecasted exposure levels. However, any deployment must be accompanied by rigorous validation against local monitoring stations, transparent governance structures, and clear documentation of data provenance to satisfy regulatory and ethical standards.

In practice, English‑language organisations that wish to adopt these forecasts should first verify that the underlying model has been independently benchmarked against national monitoring networks, confirm that the data pipelines respect privacy and data‑sharing agreements, and ensure that the provenance of both training and inference data is fully documented. Only after these checks can they responsibly integrate the predictions into operational decision‑making processes.