University of Manchester uses NVIDIA Earth-2 to forecast air pollution across the UK

Overview

Manchester physicists applied NVIDIA Earth-2 generative models (CorrDiff, later StormCast) to air quality forecasting. They generated training data from existing chemistry-climate simulations and trained CorrDiff on the Isambard-AI supercomputer using a year of hourly UK pollution data, producing a UK-wide model at 2-3 square km resolution. Training took two days on a single eight-GPU node, and the workflow also runs on an NVIDIA DGX Spark desktop system for inference and smaller training runs.

The challenge

Computing air quality with traditional chemistry-based models is expensive, which limits how detailed they can be and how regularly they can be run.

The solution

The team generated training data from existing chemistry-climate simulations, then trained NVIDIA Earth-2 CorrDiff, a generative downscaling model, on the Isambard-AI supercomputer, and has since added Earth-2 StormCast for time-dependent forecasts that use air quality observations.

Reported business value

Training the UK-wide pollution model took two days on a single eight-GPU node, and the same workflow runs on an NVIDIA DGX Spark desktop system for inference and smaller training runs. Applications such as proactive air quality alerts for healthcare organisations are described as envisioned.