University of Manchester uses NVIDIA Earth-2 to forecast UK air pollution, training models on Isambard-AI and running on DGX Spark for faster, detailed…
Air pollution contributes to an estimated 30,000 deaths in the U.K. alone last year.
The model was trained on Isambard-AI, the U.K.'s national AI supercomputer in Bristol, in just two days.
The U.K.-wide pollution model operates at a resolution of 2-3 square kilometers.
The workflow also runs on the NVIDIA DGX Spark desktop AI system for inference and smaller training runs.
Air pollution is a serious public health risk, contributing to an estimated 30,000 deaths in the U.K. last year. Traditional chemistry-based models are expensive and limit how detailed and frequent forecasts can be. David Topping, a professor at the University of Manchester, saw that NVIDIA Earth-2's generative AI frameworks had solved a similar problem for weather forecasting and asked whether they could work for pollution fields. The biggest challenge, he noted, is the compute required once chemistry is added to weather models, making them very slow. The team decided to test generative frameworks for pollution fields.
Working with NVIDIA's Earth-2 team, researchers generated training data from existing chemistry-climate simulations and trained Earth-2 CorrDiff, a generative downscaling model, on Isambard-AI, the U.K.'s national AI supercomputer in Bristol. The model worked on the first attempt. They later added Earth-2 StormCast for time-dependent forecasts using air quality observations. The training took just two days on a single eight-GPU node, and the workflow also runs on the NVIDIA DGX Spark desktop system. Topping now has a DGX Spark in his office for retraining models, noting that a few thousand dollars can get researchers started.
The U.K.-wide pollution model can simulate future scenarios, such as the impact of different government policy changes. It also offers proactive air quality insights for healthcare organizations, potentially alerting patients with asthma about high pollution in their area. The team is exploring pairing the model with edge AI devices for real-time data ingestion and decision-making, such as during wildfires. Niall Robinson of NVIDIA highlighted that the model's ability to train in two days and run on a desk changes who can do this science and how quickly.
The team plans to release open source training data and workflows so similar models can be trained for other countries and regions. Topping hopes every country and major city with a small burst of supercomputer AI time can produce their own detailed pollution models. Looking five years ahead, he envisions an agentic interface where a clinician or government agency asks a question and a chain of models delivers an answer. For example, someone could ask the pollution model running on DGX Spark about tomorrow's pollution in a specific neighborhood, with the system handling the rest.
NVIDIA Earth-2 is a family of open AI models and tools originally developed for climate and weather forecasting. Researchers at the University of Manchester adapted its generative frameworks, specifically CorrDiff and StormCast, to forecast air pollution across the U.K. The models were trained on chemistry-climate simulation data and can generate detailed pollution forecasts at a resolution of 2-3 square kilometers.
The Earth-2 CorrDiff model was trained in just two days on a single eight-GPU node of Isambard-AI, the U.K.'s national AI supercomputer. The same generative pollution workflow can also run on the NVIDIA DGX Spark desktop AI system for inference and smaller training runs, making the technology accessible to more researchers and organizations.