UK researchers train NVIDIA AI to forecast air pollution, plan open release

Forecasting air pollution is expensive. Adding chemistry to a weather model slows it sharply, limiting detail and how often forecasts run. A team at the University of Manchester says generative AI built for weather can do much of that work far more cheaply.[1]

NVIDIA's 16 September account describes how the researchers retrained its open Earth-2 models to map pollution across the UK. The account is NVIDIA's own, and its claims have not been independently tested. The team's next step matters most for Malaysia: it wants to hand the method to other countries.[1]

Professor David Topping of Manchester's department of Earth and environmental science asked whether NVIDIA's generative weather tools could work for pollution. His team generated training data from existing chemistry-climate simulations and trained Earth-2 CorrDiff, a model that sharpens coarse forecasts into fine detail. NVIDIA says it worked on the first attempt.[1]

Using a year of simulated hourly UK pollution data, the team produced a UK-wide model at a resolution of 2 to 3 square kilometres. Training took two days on a single eight-GPU node of Isambard-AI, the UK's national AI supercomputer in Bristol. The team later added Earth-2 StormCast, which makes time-based forecasts that draw directly on air quality observations.[1]

NVIDIA says the same workflow also runs for forecasting and smaller training jobs on its DGX Spark, a desktop AI computer, and that Topping now retrains models on one in his office. "You can now invest a few thousand dollars to get started developing powerful AI models," he said. No price for a full working setup is given.[1]

Topping sees uses beyond research. Health services could warn patients with asthma that pollution will be high in their area tomorrow or next week, and the model could be paired with live data from sensors to guide decisions during events such as a wildfire. These are ambitions rather than services that exist today.[1]

The team plans to release its training data and workflows as open source so that similar models can be trained for other countries and regions. "Our aim is to offer this workflow to the entire world," Topping said, adding that any country or major city with a small burst of supercomputer time could build a detailed model from its own local data.[1]

No release date is given, and the post does not say which licence the material will carry. The UK model was also trained on existing British chemistry-climate simulations. A country without comparable simulations and observations of its own would need to produce them first, which may prove harder than buying the hardware.[1]

The ASEAN Specialised Meteorological Centre's outlook, updated on 2 September, says strong El Niño conditions and below-normal rainfall could keep hotspots and smoke haze escalated across southern ASEAN, with a heightened risk of transboundary haze until the October shift out of the southwest monsoon brings some relief.[2]

On 2 September the Department of Environment reported hazardous readings in Kuching, at 312, and Serian, at 329, a day after satellites counted 980 hotspots in Kalimantan. Under the national haze plan, outdoor school activities stop when the Air Pollutant Index passes 100 and schools close above 200. On 20 September the department said it had stepped up patrols of peatland and landfill sites.[3][4]