AI Air Pollution Model Predicts UK Air Quality From Supercomputer to Desktop

Air pollution is a major public health risk. In the UK alone, it was linked to an estimated 30,000 deaths last year, according to the Royal College of Physicians.

Although data-driven insights can help researchers understand air quality, traditional chemistry-based models are expensive to run. This limits how detailed the models can be and how frequently they can produce forecasts.

David Topping, professor in the School of Earth and Environmental Sciences at the University of Manchester, explored whether NVIDIA Earth-2—a collection of open AI models and tools for weather and climate forecasting—could help solve these challenges in air pollution modeling.

“The biggest challenge is the computing required to predict air quality,” Topping said. “Incorporating chemistry into a climate model slows it down a lot, so I said, why not use a generative framework developed by NVIDIA that applies climate and weather to the pollution field?”

AI Generates Detailed UK Air Pollution Forecasts

Working with the NVIDIA Earth-2 team, Topping and his colleagues generated training data from existing chemical climate simulations. They then trained Earth-2 CorrDiff, a generative downscaling model, on Isambard-AI, the UK’s national AI supercomputer in Bristol.

The model worked on the first try.

The team then added Earth-2 StormCast, a model that supports time-dependent forecasting using air quality observations directly. They demonstrated a test training and inference workflow on the NVIDIA DGX Spark, a personal AI supercomputer.

“Understanding the effects of environmental stressors in the air we breathe is essential to improving human health,” Topping said. “Our UK-wide pollution model allows us to model potential future scenarios, including predicting what will happen if various government policy changes related to pollution are implemented.”

Video credit: Bristol Center for Supercomputing (BriCS) at the University of Bristol

Using AI to Support Health and Emergency Response

One potential application is proactive air quality information for healthcare facilities. Topping envisions a scenario in which local or national health services could contact patients with symptoms such as asthma and warn them that air pollution in their area is likely to be high tomorrow or the following week.

The team is also investigating how air pollution models could be combined with data from edge AI devices. This could enable real-time air quality monitoring and faster decision-making, including during wildfires.

“The fact that we can now train this model on Isambard-AI in two days and run it on DGX Spark on our desks changes how quickly anyone can do this science,” said Niall Robinson, developer relations manager for weather and climate at NVIDIA. “We are just at the beginning of what these open workflows can do globally.”

“The ability to switch from one NVIDIA framework to another was really impressive,” said Hao Zhang, a PhD student at the University of Manchester who trained StormCast on Isambard-AI. “We are just beginning to explore how these frameworks can be used in different ways to model complex pollution fields.”

From a National Supercomputer to a Desktop AI System

To retrain the Earth-2 air pollution model, Topping and his team used a year’s worth of UK pollution data simulated at hourly intervals. The data produced a detailed model of pollution across the entire UK at a resolution of 2 to 3 square kilometers.

Running on a single eight-GPU node of Isambard-AI, the process took just two days. Isambard-AI is the UK’s most powerful AI supercomputer, with 5,448 NVIDIA GH200 Grace Hopper superchips delivering 21 exaflops of AI performance.

“Earth-2 CorrDiff demonstrated incredibly efficient use of world-class NVIDIA hardware within Isambard-AI,” said Simon McIntosh-Smith, director of the Bristol Supercomputing Center. He holds a PhD from the University of Bristol and is a co-founder of Isambard-AI. “For climate-based projects, it is appropriate that GPU usage is relatively short and less supercomputer power is required to run the workload.”

In addition to analyzing air pollution over the past year, the model can help predict future pollution scenarios across the UK. The research team plans to increase its resolution further by incorporating additional open data, helping researchers understand air pollution at street level.

The same generative pollution workflow also runs on the DGX Spark desktop AI system, powered by the NVIDIA GB10 Grace Blackwell superchip. The system supports inference and small-scale training runs, and Topping is currently implementing it in his office to retrain the model.

“You can now invest a few thousand dollars and start developing powerful AI models,” Topping said.

Open Science Could Expand Air Quality Modeling Worldwide

The team plans to release open-source training data and a workflow for the pollution model. This would allow other countries and regions to train similar models using their own local data.

“Our goal is to offer this workflow to the entire world,” Topping said. “We hope that every country and every major city in the world with a little burst of supercomputer AI time will be able to create their own detailed pollution models using their own local data.”

Looking five years ahead, Topping envisions an even simpler interface: an agent that allows a clinician or government agency to ask a question while a set of models manages the rest of the process.

“Improved open access to air quality observations would allow someone to ask the pollution models running on DGX Spark: What will the pollution be like in this region tomorrow?” he said. “And the whole chain of interactions leads to the science-based answers that these frameworks represent.”

Learn more about NVIDIA Earth-2, NVIDIA’s climate and weather AI platform.