AIFS Blog

Abstract blue waveform lines on a dark background.

AIFL: advancing global streamflow prediction with operational machine learning

A new AI model, AIFL, improves global flood forecasting by learning from both historical data and real-world weather forecasts, delivering reliable river flow predictions under operational conditions.
Fluorescent coloured map banner

HourGlass: filling in the gaps between forecasts

HourGlass uses AI to reconstruct the atmosphere between six-hourly forecasts, producing coherent hourly forecasts.
Orange-tinted city skyline at sunset during hot, hazy conditions, with dense urban buildings beneath a glowing sun and layers of atmospheric haze stretching across the horizon.

Feeling the heat: how the AIFS learns the urban heat island

Cities stay warmer than their rural surroundings. By incorporating urban information, a prototype AIFS learns a realistic urban heat signal, paving the way for improved urban forecasting.
Satellite image of a hurricane

Improving tropical cyclone intensity forecasts with a simple AI correction

Tropical cyclones (TCs) are among the most destructive weather systems. Improving forecasts of their track and intensity is critical for helping communities to prepare for severe winds, storm surge and flooding.

Forecasting TCs has been one of the more demanding tests for machine-learning weather models.
Abstract illustration of a machine learning neural network, with interconnected nodes and data flows in blue, pink and orange converging across a dark background.

Building shared foundations for machine-learned weather prediction

The ECMWF Machine Learning Project is helping transform machine-learning research into operational forecasting through coordinated collaboration on models, data and infrastructure.
Global gridded map showing spatial variations in blue and orange tones.

Adapting the AIFS for 50r1

How AIFS training was adapted for IFS Cycle 50r1.
Abstract network graphic with glowing nodes connected by lines on a blue background, representing data or AI connections.

Farewell to the external AI models

The upgrade to IFS Cycle 50r1 highlights limitations in externally developed ML models, prompting a discontinuation of their real-time use.
Anemoi image showing wind wheels and a cloud

Anemoi: a European framework for operational AI in weather and climate

Anemoi, developed collaboratively by ECMWF and national meteorological services, provides the shared foundations needed to move AI from research into real-world operations.
Weather map showing a deep low‑pressure system with coloured wind or precipitation fields and contour lines indicating pressure

Severe storms in the age of machine learning

Comparison of ECMWF’s physics-based and machine learning forecasts for two major storms highlights strong large-scale performance but ongoing challenges in capturing localised wind extremes and storm structure.
Cyclone Mocha from Space

Evaluating training strategies in tropical cyclone forecasts

This blog examines how training strategies affect tropical cyclone forecasts in machine-learning weather models.
Purple map of globe with patches of red, light pink and blue

Does better mean better forecasts for everyone?

Are improvements from the ML-based AIFS system distributed evenly worldwide or do some regions benefit more than others?
A very snowy landscape with a field and tress in the background all covered in snow. The sky above is clear blue

Let it snow! How machine learning will forecast snow in the AIFS

AIFS v2 adds snow fields, improving snow forecasts and advancing ECMWF’s ML-based Earth system modelling across land, oceans, hydrology, sea ice and waves.
Dark blue banner image with circular network showing connections. Over the top are red lines and colourful dots

Introducing the Anemoi training-ready version of ERA5

The Anemoi training-ready version of ERA5 is now openly available. This dataset provides a ready-to-use machine-learning-optimised subset of ERA5 data enabling efficient training of models.
Card format image showing extract from significant wave height forecast

Representing ocean wind waves in ECMWF's AIFS

We present data-driven ocean wave modelling for the AIFS, introducing two methods for integrating waves alongside the atmosphere. This improves the wave forecast accuracy and advances machine-learning-based Earth system modelling.
AIFS blog card 2m temperature verification

Verifying 2 m temperature forecasts in wintertime anticyclonic conditions

We provide a conditional verification of two-metre temperature forecasts from the AIFS Single (ML-based deterministic forecasts) and IFS Control forecasts during the winter of 2024/25, focusing on synoptic weather patterns to better understand when and where differences in forecast skill emerge.
Anemoi webinar graphics

Discover Anemoi: Kicking off our 2025 machine learning training

In January, we launched our 2025 machine learning training webinars called 'Discover Anemoi'. The webinar series introduces the features and usage of the Anemoi AI weather forecasting framework.
World map with graph used in graph neural network processors in Anemoi

A year in ML for weather forecasting

A tour of the year past and some previews of the year to come.
Graphic of a map

First AIFS model weights are now open

We highlight that the first AIFS model weights are now available and explain how you can use them yourself.
Excerpt from Aurora ML-based forecast chart

Accuracy versus activity

Forecast performance is multifaceted. We discuss how accuracy and activity can be ‘competing’ attributes for ML-based deterministic forecasts.