University of Manchester Uses NVIDIA Earth-2 for UK Air Quality Predictions
Using NVIDIA Earth-2, University of Manchester researchers deliver air pollution forecasts at unprecedented speed and resolution.
The University of Manchester is leveraging NVIDIA’s cutting-edge Earth-2 AI models to revolutionize air pollution forecasting across the UK. By integrating NVIDIA’s generative AI technology, researchers are now able to generate high-resolution, time-sensitive air quality predictions, potentially transforming how public health decisions are made.
Air pollution is a critical health issue in the UK, linked to an estimated 30,000 deaths annually, according to the Royal College of Physicians. Yet, traditional chemistry-based forecasting models are computationally expensive, limiting their frequency and detail. Using NVIDIA’s Earth-2 CorrDiff and StormCast models, Manchester scientists David Topping and Hao Zhang have overcome these challenges. The models produce rapid, high-resolution forecasts that can integrate real-time air quality observations.
The workflow used NVIDIA's Isambard-AI supercomputer in Bristol—currently the UK’s most powerful AI setup. Training the model took just two days on Isambard-AI, which is powered by 5,448 NVIDIA GH200 Grace Hopper Superchips delivering 21 exaflops of AI performance. The resulting model covers air pollution at a granular 2-3 square kilometer resolution across the UK. Remarkably, these workflows can now run on a desktop NVIDIA DGX Spark system, democratizing access to this technology.
“The ability to run these models on a DGX Spark system—a machine costing just a few thousand dollars—changes the accessibility of this science,” said Topping. “This isn’t just about national supercomputers anymore; an individual researcher can now iterate and experiment in their own office.”
The implications are significant. Public health agencies could use the data to proactively inform individuals vulnerable to respiratory issues, such as asthma patients, about upcoming high-pollution days. The team is also exploring integration with edge AI devices to enable real-time decision-making during events like wildfires.
Air pollution forecasting is the latest application in NVIDIA's Earth-2 platform, which has disrupted climate and weather modeling with its generative AI approach. Announced in January 2026, the broader Earth-2 family includes tools for global climate simulation and medium-range weather forecasting, emphasizing speed, resolution, and energy efficiency. NVIDIA claims Earth-2 models are up to 1,000x faster and 3,000x more energy-efficient compared to traditional numerical models.
The University of Manchester plans to release its training data and workflows as open source, enabling other countries and cities to develop localized pollution models. Topping envisions a future where these models become even more accessible, with users relying on agentic systems to ask questions and receive immediate insights. For example, a clinician could query the system for neighborhood-specific pollution forecasts, with the models handling all the complexity in the background.
NVIDIA’s ongoing advancements in AI weather and climate modeling are rapidly broadening their utility beyond research labs. For traders and investors, this highlights NVIDIA’s continued dominance in the AI hardware and software space, solidifying its position as the go-to provider for high-performance AI applications. With a market cap of $5.15 trillion and a stock price of $212.17 (as of September 16, 2026), NVIDIA remains at the forefront of AI-driven innovation, creating opportunities in sectors from public health to environmental science.
Asanat Analysis — Why it matters
This deployment signals a maturing shift in AI infrastructure: specialized climate/weather modeling, traditionally reserved for national supercomputing centers, is being democratized to edge hardware. The 2-3km resolution air quality forecasting at 1,000x speed gains over classical chemistry models represents a category change in real-time public health decision-making capability. What matters here isn't the novelty of generative AI for climate (Earth-2 launched in January 2026), but the operational accessibility—researchers can now iterate on local models using <$10k hardware rather than competing for supercomputer clusters.
This directly impacts regulatory compliance and urban planning workflows. UK air quality management has historically relied on sparse monitoring networks and batch-processed forecasts. High-resolution, low-latency predictions enable proactive warnings to vulnerable populations (asthma, COPD patients) and real-time pollution response during pollution events or wildfires. The open-sourcing of training data and workflows creates network effects: other jurisdictions can rapidly customize models, compressing the adoption curve for AI-native environmental monitoring.
Strategically, this reinforces NVIDIA's moat in enterprise AI beyond gaming/training—they're embedding themselves as the foundation layer for climate-tech and environmental intelligence, a growing regulatory requirement globally. The story also indirectly validates the energy-efficiency gains of generative approaches (3,000x efficiency claim), relevant as data centers face mounting power constraints.