Weather forecasting is the obvious application of Aurora, but it is not the only one. The Microsoft team explicitly framed Aurora as an atmospheric foundation model, meaning the same backbone is fine-tuned into multiple task-specific heads. Two of the most interesting heads are air quality prediction and tropical cyclone tracking. This post walks through what each one does and how to use them through SciRouter's climate lab.
Air quality prediction with Aurora
Air quality forecasting is a stack of hard problems. You need to know how winds are transporting air masses, how pollutants are being emitted from sources, how they are chemically transforming in the atmosphere, and how boundary-layer mixing is distributing them vertically. Traditional chemical transport models like GEOS-Chem solve all of this with physics and chemistry, and they are excellent — but they are slow.
Aurora's air quality head takes the shared atmospheric representation and maps it directly to pollutant concentrations. The model has learned the relationships between atmospheric dynamics and pollutant transport from training data, so it can produce a global air quality forecast in a fraction of the time.
What it predicts
- Particulate matter (PM2.5, PM10). The pollutants that matter most for human health, on a global grid.
- Nitrogen dioxide (NO2). A marker for combustion sources and a precursor for ozone and secondary particulate matter.
- Sulfur dioxide (SO2). Important for industrial emission tracking and secondary aerosol formation.
- Ozone (O3). The ground-level photochemical pollutant that matters for respiratory health.
- Carbon monoxide (CO). A long-lived marker for combustion and long-range transport.
How to use it
You send a starting atmospheric state and the forecast horizon to the Aurora air quality endpoint through SciRouter. The response is the predicted concentrations of each pollutant on a global grid at the requested lead time. If you pass several horizons in one call, you get a time series.
Typical applications include real-time dashboards that show air quality trajectories for a city or region, what-if analysis of how wind-shift scenarios affect pollution transport, and rapid scoping for emergency-response planning. Aurora's speed is what makes all of these interactive.
Tropical cyclone tracking with Aurora
Tropical cyclones are one of the highest-leverage forecasting problems in the world. A better forecast of where a storm will go in the next 72 hours directly translates to better evacuation decisions and fewer deaths. The operational forecasts from agencies like NOAA and ECMWF are the gold standard, and they should stay the primary source for life-safety decisions.
Aurora's cyclone head adds something new: the ability to generate large ensembles of track forecasts in seconds. Because Aurora is fast, you can perturb the initial atmospheric state many ways and watch how the storm track spreads, giving you a probabilistic view that used to be available only to the largest forecasting centers.
What it predicts
- Storm position. Predicted latitude and longitude of the storm center at future lead times.
- Intensity indicators. Depending on the model variant, associated intensity measures such as minimum central pressure and maximum wind speed.
- Uncertainty. When run as an ensemble, a cone of plausible tracks, not a single deterministic line.
How to use it
Send a recent atmospheric state that already contains an active cyclone. Aurora does not invent storms out of clear skies — it forecasts forward from the state you give it. If you want to track a specific storm, pull a fresh state from ERA5 or another source and feed it to the endpoint.
For ensemble tracking, run the same storm state multiple times with small perturbations in wind, temperature, or humidity. Aurora's speed means a 50-member ensemble takes less time than one physics-based forecast would.
Combining the two tasks
The interesting pattern is chaining the cyclone and air quality heads together. A landfalling storm redistributes air quality — wildfires push smoke around, industrial zones experience sudden changes in pollutant advection, and coastal communities see PM2.5 spikes or drops depending on wind direction. Aurora lets you reason about both in the same workflow.
- Pull a current atmospheric state.
- Call the cyclone head to forecast the storm track.
- For each time step, call the air quality head to forecast pollutant fields.
- Render both layers on the same map to see how storm motion and air quality interact.
This kind of composite forecast is exactly what foundation-model framing was designed for. One backbone, many heads, all sharing representation.
Limits to keep in mind
- Out-of-distribution behavior. Aurora is trained on historical data. If atmospheric conditions are truly unprecedented, the model's confidence may be miscalibrated.
- Local effects. A global model cannot resolve every city block. For hyperlocal air quality questions you still want urban-scale chemistry models.
- Emissions changes. Sudden emission changes — say, a major industrial shutdown or a wildfire event — may not be captured if the initial state does not reflect them.
- No attribution. Aurora tells you what it thinks will happen. It does not tell you why in the physical-mechanistic sense. For that you want a coupled chemistry-climate model.
Bottom line
Aurora's air quality and tropical cyclone heads are what make the foundation-model framing real. A single backbone trained on the atmosphere can now power forecasts that used to require entirely separate modeling stacks, at speeds that open new applications. Hooked up through SciRouter's climate lab, it is a practical tool for anyone building climate, health, or emergency-response products.