The Inference Primerinductive inference · information · models

Case study

How machine-learning weather forecasts work

A weather forecast asks how the atmosphere might change over an interval of time. In a machine-learning forecast, the inference draws on patterns learned from how weather has developed in the past.

The weather at the start

On 25 February 2025, the European Centre for Medium-Range Weather Forecasts (ECMWF) put AIFS Single into operation alongside its physics-based Integrated Forecasting System (IFS). AIFS Single was the first operational version of ECMWF's Artificial Intelligence Forecasting System. The two systems used the same estimate of the atmosphere as the starting point for a forecast. That estimate is known as the initial conditions.

At launch, ECMWF formed that estimate by combining an earlier short-term forecast with around 60 million observations checked for quality. Satellites supplied some of those observations, alongside measurements from aircraft, ships, sea buoys and other measurement stations. ECMWF fed the starting estimate into AIFS Single every six hours. From each such estimate, AIFS Single forecasts conditions over the coming days.

One forecast or a range of possible futures

AIFS Single makes one forecast at a time, a form known as a deterministic forecast. On 1 July 2025, ECMWF put AIFS ENS into operation alongside the physics-based IFS. AIFS ENS produces an ensemble, a collection of 51 forecasts at a time with slight variations that describe different possible courses for the weather. For AIFS ENS, machine learning projects the future from initial conditions prepared by a physics-based process called data assimilation.

An ensemble can also pool forecasts from different models. In December 2025, the National Oceanic and Atmospheric Administration (NOAA) announced its operational HGEFS, a 62-member ensemble that combines 31 forecasts from the physics-based GEFS with 31 from the AI-based AIGEFS.

A physical limit on rainfall

An amount of precipitation can be zero or positive. A June 2026 study of AIFS Single 1.1.0 in Geoscientific Model Development described how ECMWF built that limit into the model by adding a layer that bounds its output. The layer also keeps related outputs consistent within the precipitation forecasts and within the cloud-cover forecasts. The revision also expanded the training data, changed the weights used to measure errors during training and added surface and atmospheric variables.

The study found that the lower limit resolved an ambiguity about how the model should adjust at the boundary of zero precipitation during training with mean squared error. The study linked that change to less drizzle bias and attributed the largest improvements in precipitation forecasts to the output bounds. In a controlled comparison, the expansion of the training data was the main source of improved forecasts in the upper air.

What the forecast was checked against

The December 2024 Nature study of GenCast described a model that produces worldwide 15-day ensemble forecasts at a resolution of 0.25 degrees. During development, the study trained GenCast on data from 1979 through 2017 and used 2018 for validation. Before the test, it fixed the model and training choices and retrained GenCast on the full 1979–2018 span. It then evaluated the results on 2019 weather.

Against ECMWF's operational ensemble system ENS, the study reported better probabilistic forecast skill for GenCast on 97.2% of the 1,320 targets evaluated. Some of the weather variables depend on one another, so the scorecard includes related forecast questions. The study also reported improvements in predictions of extreme weather, tropical cyclone tracks and wind power production. The study excluded precipitation from its main results because the team lacked confidence in the quality of the ERA5 precipitation data.