University of Manchester Uses NVIDIA Earth-2 to Forecast Air Pollution Across the UK
The University of Manchester is using NVIDIA’s Earth-2 AI technology to develop high-resolution forecasts of air pollution across the United Kingdom, potentially making detailed pollution modeling faster and more accessible to researchers, healthcare organizations and policymakers.
The project, led by David Topping, a professor in the university’s Department of Earth and Environmental Science, uses generative AI models to tackle a major challenge in air-quality forecasting: the enormous computing requirements of traditional chemistry-based atmospheric models.
Air pollution remains a significant public-health concern in the UK. Government sources estimate that outdoor air pollution is associated with roughly 29,000 to 43,000 deaths each year, while exposure can contribute to respiratory and cardiovascular disease and worsen conditions such as asthma.
How NVIDIA Earth-2 Is Being Used for Air Pollution
Topping and his colleagues began exploring whether generative AI techniques developed for weather and climate modeling could also be applied to pollution.
The researchers worked with NVIDIA’s Earth-2 team and generated training data from existing chemistry-climate simulations. They then trained Earth-2 CorrDiff, a generative downscaling model, using Isambard-AI, the UK’s national AI supercomputer based in Bristol.
According to the project team, the initial model produced results on its first training attempt.
The researchers subsequently incorporated Earth-2 StormCast, allowing the system to work with time-dependent forecasting and air-quality observations.
NVIDIA has described related University of Manchester work as PolluGen, a high-resolution pollution-dispersion model built using NVIDIA CorrDiff and Earth-2 Studio. The company says the system can use regional health and other data to help inform citizens and policymakers about air-quality impacts.
Two Days of Training on Isambard-AI
The team used a year’s worth of simulated UK pollution data, generated at hourly intervals, to build its initial nationwide model.
The resulting system can model pollution across the UK at approximately 2–3 square kilometer resolution, according to the project information.
Training was carried out on a single eight-GPU node on Isambard-AI and reportedly took about two days.
Isambard-AI is equipped with 5,448 NVIDIA GH200 Grace Hopper Superchips and is designed for large-scale AI workloads. The Bristol Centre for Supercomputing says the system delivers 21 exaflops of AI performance.
The researchers say the relatively short training time demonstrates how generative AI could reduce some of the computational barriers associated with detailed environmental modeling.
NVIDIA’s Earth-2 family is designed to provide open AI models and tools for weather and climate applications. Its CorrDiff technology is intended to generate higher-resolution regional fields from coarser-scale data.
From a Supercomputer to a Desktop AI System
One of the more notable aspects of the project is that the workflow does not remain confined to a national supercomputer.
The researchers have also demonstrated inference and smaller training workloads on NVIDIA’s DGX Spark, a desktop AI system powered by the GB10 Grace Blackwell Superchip.
That could change how researchers experiment with pollution models.
Instead of requiring continuous access to a large supercomputer, researchers could potentially retrain or test models locally after the initial large-scale development work.
The University of Manchester team says it is now using a DGX Spark system to retrain models.
Could AI Predict Tomorrow’s Pollution?
The researchers see several possible applications beyond creating historical pollution maps.
One is advance warning for people who are particularly vulnerable to poor air quality.
For example, a healthcare organization could potentially use future pollution forecasts to alert patients with asthma or other respiratory conditions when pollution levels are expected to rise in their area.
That concept is still an envisioned application rather than an established nationwide healthcare service.
The team is also investigating whether pollution models could eventually combine information from edge AI devices and real-time air-quality sensors.
Such a system could be useful during rapidly changing events, including wildfires, where pollution conditions can change significantly over short periods.
Researchers Want Street-Level Pollution Forecasts
The current model provides UK-wide coverage at kilometer-scale resolution, but the researchers want to make it considerably more detailed.
The next step involves incorporating additional open datasets to increase the model’s resolution and potentially provide information closer to street level.
That could make the technology more useful for studying how pollution varies between neighborhoods, roads and other localized environments.
The project could also allow researchers to test hypothetical scenarios.
Topping said the model could be used to explore what might happen if governments introduced different pollution-related policies, providing a way to examine potential future air-quality conditions.
The UK’s existing air-quality framework already includes targets for reducing fine particulate matter, including a target for annual mean PM2.5 concentrations of 10 micrograms per cubic meter or lower by 2040.
The Bigger Goal: Open Pollution Models
The University of Manchester team plans to release training data and workflows so researchers in other countries and regions can potentially build similar models using local information.
The idea is to create a reusable framework rather than a system limited to the UK.
Topping envisions cities and countries being able to combine local pollution observations with the open AI workflow and generate their own detailed forecasting systems.
The longer-term vision is even more automated.
Researchers imagine a future where a clinician, government agency or other user could simply ask an AI system something like what pollution is expected in a particular neighborhood the following day.
Behind the scenes, a chain of scientific models could process observations, generate a forecast and return an answer grounded in the underlying environmental data.
That remains a future direction rather than a deployed public service today.
For now, the University of Manchester project demonstrates how generative AI originally developed for weather and climate applications can be adapted to another difficult forecasting problem: understanding where air pollution is likely to appear, how it may change and how quickly researchers can generate detailed scenarios.
