How Modern Weather Forecasting Works: From Satellites to Supercomputers

Every weather forecast, whether it is a seven-day outlook on a phone app or a hurricane track issued by a national agency, starts the same way: with observations of the atmosphere as it is right now. The European Centre for Medium-Range Weather Forecasts (ECMWF) says it receives around 800 million observations a day, mostly from about 90 satellite instruments monitoring the atmosphere, ocean, land, snow and sea ice; after quality control, roughly 60 million of those observations are actually used in its forecasting system.

Turning observations into a forecast: data assimilation

Raw observations do not cover every point on Earth, so forecasters cannot simply read off the current weather from them. Instead, agencies use a technique called data assimilation to combine incomplete observations with a short-range forecast from the previous run, producing a complete, physically consistent picture of the atmosphere called an analysis. ECMWF uses a method called four-dimensional variational assimilation (4D-Var), which it describes as combining 'a first guess of the current weather with new evidence from observations' across both space and time. The UK's Met Office describes the same step more simply: data assimilation 'fills in gaps where direct observations are missing... by blending available data with model predictions.'

Among the observation types prioritised by ECMWF are temperature and humidity readings from microwave and infrared satellite instruments, wind estimates derived from tracking clouds and from radar, lidar and scatterometer instruments, and a technique called 'all-sky' assimilation that uses satellite data from cloudy and rainy areas instead of discarding it.

Running the models: physics equations on supercomputers

Once the analysis is built, a numerical weather prediction model takes over. These models divide the atmosphere into a three-dimensional grid and solve equations governing motion, energy, mass and moisture at each grid point, then step the whole system forward in time. The US Global Forecast System (GFS), run by NOAA's National Centers for Environmental Prediction, uses a base horizontal resolution of about 28 kilometres (18 miles) between grid points for its first week of output, coarsening to about 70 kilometres (44 miles) for forecasts beyond seven days, and couples separate models of the atmosphere, ocean, land/soil and sea ice. It produces a full forecast four times a day, out to 16 days.

ECMWF's equivalent system, the Integrated Forecasting System (IFS), was upgraded in November 2024 under a release called Cycle 49r1. According to ECMWF, that upgrade assimilated two-metre temperature observations for the first time, revised how the model represents uncertainty in its physics, and matched the ocean-wave model's grid to the atmospheric model's resolution of roughly 9 kilometres for medium-range forecasts. ECMWF reported measurable gains from the upgrade, including about a 0.5% improvement in forecast scores near Antarctica through day four, driven partly by new microwave data over sea ice.

Ensemble forecasting: predicting the uncertainty itself

A single model run from a single analysis gives one version of the future, but the atmosphere is what scientists call a chaotic system: according to the Met Office, 'very small errors in its initial state can lead to large errors in the forecast.' Because observations can never capture the atmosphere's true state perfectly, agencies run the same model dozens of times from slightly different, equally plausible starting conditions. This collection of parallel forecasts is called an ensemble, and each individual run is an 'ensemble member.'

Why forecasts lose confidence further ahead

ECMWF has reported that its November 2024 Cycle 49r1 upgrade produced measurable gains in forecast accuracy, including about a 0.5% improvement in forecast scores near Antarctica through day four, which it attributed partly to new microwave satellite data over sea ice. More broadly, because small errors in the starting conditions of a chaotic system grow over time, agencies also issue separate monthly and seasonal outlooks, based on different statistical methods, once day-by-day forecasts lose their usefulness.

The rise of AI-based forecasting

Alongside these physics-based models, national agencies have begun running machine-learning systems trained on decades of past weather data. NOAA announced in December 2025 that it had made three such systems operational: AIGFS (an AI version of the Global Forecast System), AIGEFS (an AI ensemble system) and HGEFS, a 62-member ensemble that blends AI and physics-based forecasts. According to NOAA, AIGFS can produce a 16-day global forecast in about 40 minutes using roughly 0.3% of the computing resources the traditional GFS requires, and AIGEFS needs about 9% of the resources of the standard ensemble. NOAA Administrator Neil Jacobs described the systems as 'a significant leap forward in American weather model innovation,' citing improved accuracy for large-scale patterns and hurricane track forecasts, though NOAA also acknowledged that AIGFS currently shows degraded performance forecasting hurricane intensity, an area officials said still needs further development.

From model output to the forecast you see

Model data is not usually the final product. The Met Office says that after the supercomputer runs are complete, human forecasters interpret the model output, factor in local terrain and conditions the model's resolution cannot fully capture, and adjust the forecast before it is published, combining human judgement with the computational output to produce a usable forecast.