A team at the University of Manchester has used Nvidia’s Earth-2 models to produce air quality forecasts covering the UK at a resolution of two to three square kilometres, according to an account published on 15 September. The group, led by Professor David Topping with doctoral researcher Hao Zhang, trained the Earth-2 CorrDiff model on historical pollution data and added StormCast to generate time-dependent forecasts that take in actual air quality observations.
The figure worth noting is the training cost. The model was trained in about two days on a single eight-GPU node of Isambard-AI, the UK’s national AI supercomputer in Bristol, which is built from 5,448 Nvidia GH200 Grace Hopper superchips. One node. The team also reports running the same workflow on a desktop DGX Spark for inference and smaller retraining runs.
Air quality modelling has traditionally meant chemical transport models that are costly to run and slow to produce fine-grained output. If a comparable forecast can be trained on a single node and re-run on a desk, the binding constraint shifts from compute to data — and the team says it intends to release its training data and workflows so the approach can be repeated elsewhere.
One caveat: these figures come from Nvidia’s own account of the work rather than from a peer-reviewed paper.
Source: Nvidia.