Developing a forecasting system powered by machine learning

Machine learning (ML) is an emerging tool that will help us solve some problems in weather and climate forecasting that are difficult to address with conventional, physics-based approaches. In 2023, we built a first version of a forecasting system powered by ML and brought it to a preoperational stage.

A significant reason that our AIFS, and other experimental models, exist is because of the ERA5 reanalysis, produced by the Copernicus Climate Change Service at ECMWF. High-quality data are vital for training high-quality models, so the open and easy-to-access ERA5 database has been pivotal in driving this new generation of models.

Florian Pappenberger, Director of Forecasts and Services

ECMWF is advancing the science of prediction with physics-based systems, hybrids of physics and data-driven approaches, and fully data-driven models. All of these strands will contribute to continuing to improve forecasts for our users.

Andy Brown, Director of Research

Artificial intelligence and particularly ML have dominated headlines, and their use in weather forecasting is no exception.

We continued following our ML Roadmap, published in 2021, which targeted bringing ML into conventional forecasting workflows, including into our Integrated Forecasting System (IFS). At the beginning of the year, we began to use ML in the operational pipeline, to help monitor meteorological observations. By the end of the year, we had developed a fully ML-based forecasting system.

In 2023 our Member States approved two initiatives aimed at intensifying the development, testing and implementation of ML across weather forecasting processes, and expanding our existing applications of ML to Earth system modelling.

The first project, approved in June, looks at increasing resources within ECMWF to broaden the scope of our ML work. This will include a range of approaches, from expanding existing work to bring ML into the IFS, to developing a fully data-driven forecasting system.

Resources were also allocated for a collaboration with our Member and Co-operating States. This new ‘pilot project’ was approved by the December session of Council. Working together, we can move faster in the rapidly evolving discipline of applying ML to weather prediction.

The two projects together recognise the potential of ML and build on substantial activities in this field, both at ECMWF and within European meteorological services.

A significant reason that our AIFS, and other experimental models, exist is because of the ERA5 reanalysis, produced by the Copernicus Climate Change Service at ECMWF. High-quality data are vital for training high-quality models, so the open and easy-to-access ERA5 database has been pivotal in driving this new generation of models.

Florian Pappenberger, Director of Forecasts and Services

ECMWF is advancing the science of prediction with physics-based systems, hybrids of physics and data-driven approaches, and fully data-driven models. All of these strands will contribute to continuing to improve forecasts for our users.

Andy Brown, Director of Research

A significant output of the first project was the Artificial Intelligence Forecasting System, or AIFS, with a first version launched in October. This data-driven model is trained to go from analysis to future forecast with a system that learns physical relationships in historical weather data. Development of data-driven models, including the AIFS, will be carried out in collaboration with Member States, led by Norway and Switzerland.

To assess the effectiveness of the AIFS, we evaluated its ability to forecast Storm Ciarán. This exceptional weather event in early November 2023 was well forecast by traditional methods, allowing warnings to be issued and authorities and individuals to take effective actions. The evaluation showed that the AIFS predicted the development of the cyclonic system very well but underestimated the most severe winds. This is a known current weakness of the AIFS, which partially stems from the fact that the training data used at the time was at a low resolution. For the scientific team, this test case was a positive demonstration of its potential use as a primary forecasting tool.

On release, the AIFS was added to our open charts. During 2023, we also added graphical products from experimental models developed by Huawei, NVIDIA and Google DeepMind, and one built through a higher education collaboration led by Fudan University (China). By showcasing the models in this way, we allow the meteorological community to compare these systems, and demonstrate their strengths and weaknesses.

Towards the end of 2023, we also began work on another angle of ML: a radical project to investigate if weather forecasts can be made directly from meteorological observations, without the use of reanalysis. The year was ambitious and successful for our ML work, coming a very long way in a very short space of time, and laying solid foundations for us to expand and improve the IFS and AIFS even further in 2024.


AIFS alpha version | October 2023  
 


Architecture: graph neural networks 

Grid: reduced Gaussian grids 

Pressure levels: 13 

Resolution: ~1 degree 

Predicted parameters: wind, temperature, humidity, geopotential, surface pressure 

Training data: ERA5 reanalysis and IFS operational analysis   
 

 

Background image: extract from a 48-hour forecast for mean sea-level pressure and wind speed for 00 UTC on 2 November 2023 (Storm Ciarán) from an early, low-resolution version of ECMWF’s experimental AIFS machine learning model.