UK Scientists Leverage NVIDIA Earth-2 for Real-Time Air Pollution Predictions
Researchers at the University of Manchester used NVIDIA's Earth-2 generative AI models to forecast air pollution across the UK, overcoming the high computational cost of traditional chemistry-based models. The team trained CorrDiff on the Isambard-AI supercomputer and added StormCast for time-dependent forecasts, enabling predictions that could inform public health decisions.
The project repurposed NVIDIA's weather-focused Earth-2 frameworks for pollution modeling, training CorrDiff on a year of hourly UK pollution data to produce forecasts at 2-3 square kilometer resolution. Training required just two days on a single eight-GPU node of Isambard-AI, the UK's national AI supercomputer equipped with 5,448 GH200 Grace Hopper Superchips.
The team subsequently integrated StormCast to enable time-dependent predictions that incorporate live air quality observations. The workflow now runs on a DGX Spark desktop system, and researchers are exploring pairing the model with edge AI devices for real-time decision making during events like wildfires.
This approach could make high-resolution air quality forecasting practical for routine public health use. Healthcare organizations may proactively warn asthma patients about impending pollution spikes, and policymakers could model the effects of proposed regulations before implementation. The dramatic reduction in computational requirements — from national supercomputer to desktop — could democratize this type of environmental modeling, potentially enabling more communities worldwide to develop localized forecasting capabilities.