The Meteorology Stack: What Modern Forecasting Actually Runs On

Ask most people how weather forecasting works and they'll describe something like "supercomputers running simulations". That's true as far as it goes — but the modern operational forecasting stack is much more than a physics model. It's a pipeline of data ingest, blending, modelling, post-processing and delivery, and every stage matters.

Here's a plain-English tour of what actually runs underneath a professional weather service.

Layer 1: Observations

Everything starts with observations. There are far more of them than most people realise:

• Surface stations: ~10,000 automated weather stations globally reporting temperature, pressure, wind and humidity every minute.
• Weather radar: National networks (like the UK Met Office's) scanning precipitation every 5 minutes at multiple elevation angles.
• Satellites: Geostationary and polar-orbiting satellites providing cloud imagery, temperature soundings, and atmospheric motion vectors.
• Aircraft: Thousands of commercial flights reporting wind, temperature and turbulence (via AMDAR and similar programmes).
• Radiosondes: Balloon-launched sensors giving vertical profiles twice a day at ~800 sites worldwide.
• Ground-based remote sensing: Wind profilers, lidar, microwave radiometers, and increasingly cheap IoT sensors.

Layer 2: Data assimilation

Raw observations are messy — different formats, different quality, different timing. Data assimilation is the process of combining observations with a short-range model forecast (the "background") to produce the best possible estimate of the current atmospheric state. This is the analysis, and it's the input to the forecast model.

Modern assimilation uses 4D-Var or ensemble Kalman filter techniques, and can ingest hundreds of millions of observations per day.

Layer 3: Numerical models

The heart of the stack. Numerical weather prediction (NWP) models solve the equations of atmospheric motion on a grid. Three families dominate:

• Global models: ECMWF IFS, NOAA GFS, UKMO UM. ~9–25km resolution, out to 10–15 days.
• Regional models: Downscale global models over a specific area. Typically 1–4km resolution.
• Convection-permitting models: 500m–1.5km grids that resolve individual storm cells. This is where we operate.

"Resolution is not a vanity metric. It is the difference between a forecast you can act on and a forecast you can only react to."

Layer 4: Post-processing

Raw model output has biases. It tends to be too smooth, too coarse, and slightly off on things like wind gusts and precipitation intensity. Post-processing corrects these using statistical methods:

• Model Output Statistics (MOS): Regression against historical observations.
• Ensemble dressing: Combining the members of an ensemble forecast into a probability distribution.
• Machine learning: Increasingly, neural networks are used to correct systematic biases and blend multiple models.
• Nowcasting: Radar extrapolation for the 0–2 hour window, where NWP can't beat persistence.

Layer 5: Delivery

A forecast is only useful if it reaches the decision-maker in time and in a usable format. Delivery involves:

• Streaming infrastructure: Kafka or similar to move data between services at low latency.
• APIs: REST and WebSocket endpoints that client systems can query in real time.
• Dashboards: Purpose-built UIs for control rooms, dispatch desks and executive teams.
• Alerting: Rules engines that push notifications when thresholds are crossed.

Layer 6: Feedback

Every operational forecast generates a feedback signal. Did the storm arrive as predicted? Did the wind gust hit the threshold? Was the outage in the forecast zone? Good weather companies close the loop and feed this back into model tuning, post-processing and client dashboards.

Putting it together

At Nebula Storm, our stack looks roughly like this: observations (radar, satellite, surface, aircraft) → our streaming ingest layer → a blended multi-model analysis → a 500m convection-permitting forecast → ML-based post-processing → API and dashboard delivery → operator feedback → model tuning.

Every layer matters. Skip one and the whole thing degrades. Get them all right and you get a forecast that people actually trust with real decisions.