A team from the University of Manchester, in collaboration with the NVIDIA Earth-2 team, has developed an artificial intelligence-based model to simulate and predict air pollution across the United Kingdom, at a spatial resolution of between two and three kilometers. The project aims to reduce the high computational cost of traditional models that combine weather data with chemical interactions in the atmosphere.
The work is led by David Topping, a professor in the Department of Earth and Environmental Sciences at the University of Manchester. The team used pollution data generated from a chemical-climate simulation covering an entire year at hourly intervals, then trained the Earth-2 CorrDiff model to generate more detailed maps of pollutant distribution. According to NVIDIA, the model succeeded on its first training attempt.
From Chemical Simulation to Rapid Prediction
Traditional air-quality models require intensive computation because they incorporate atmospheric chemistry into weather models, limiting the level of detail and the number of times they can be run. The Manchester team sought to use the generative artificial intelligence frameworks developed by NVIDIA for weather and climate to handle pollution “fields” rather than calculating every interaction in the traditional way.
Training used a single node equipped with eight graphics processing units on the British Isambard-AI supercomputer in Bristol. The system contains 5,448 NVIDIA GH200 Grace Hopper Superchips, with a stated capacity of 21 exaflops for artificial intelligence workloads. Training took two days, according to the published material.
The team later added the Earth-2 StormCast model to produce time-dependent forecasts that directly use air-quality observations. The researchers also demonstrated training and inference workflows on the desktop NVIDIA DGX Spark system, powered by the GB10 Grace Blackwell chip. Topping says that having a DGX Spark in his office allows him to retrain the models and conduct smaller training runs away from the supercomputer.
Why Does This Development Matter?
The material says that air pollution contributed to approximately 30,000 deaths in the United Kingdom during the past year, making improved forecasts directly relevant to public health. Health services could use these forecasts to inform patients with conditions such as asthma about an expected rise in pollution in their area tomorrow or during the following week.
The team is also testing connections between the model and edge artificial intelligence devices that receive real-time air-quality measurements. Proposed applications include supporting rapid decisions during wildfires, as well as using the model to simulate the effects of potential changes in government pollution policies.
What Changes in Practice?
The most significant change is the shift from costly centralized runs to a model that can be run, at least for inference and some training tasks, on a desktop system. The team plans to increase the model’s resolution to the street level by integrating additional public data, and also intends to publish the training data and workflows as open source so that other countries and cities can build local models using their own data.
However, the material does not provide detailed quantitative results on forecast accuracy compared with traditional chemical models, nor has it yet demonstrated the operation of a health service or public warning system based on the model. Therefore, the promises concerning real-time forecasts and clinical applications remain goals under exploration, not services that are already available. The project’s current significance lies in demonstrating the portability of Earth-2 frameworks to the field of air pollution and lowering the barrier to accessing modeling tools, while independent validation, the quality of local data, and the timing of open-resource publication remain questions requiring answers.