A storm wreaks havoc when the forecast promised a breeze. A sunny day turns into a torrential downpour. The hoped-for snow day does not materialize. Having plans disrupted by a faulty forecast is a common experience, but artificial intelligence (AI) could be about to change that.
Meteorological agencies across Europe, the United Kingdom and the United States have been testing AI-generated forecasting programs and machine learning (ML) weather models.
These efforts are now producing measurable results. The UK's Met Office, for instance, reached a major milestone: improved training methodologies mean its experimental AI weather model, called Fastnet, now performs on par with the agency's physics-based modeling.
A similar pattern has emerged during the current heatwave over the UK and Ireland, where the traditional forecasting system used by the European Centre for Medium-Range Weather Forecasts (ECMWF) has been outperformed by its AI counterpart, the AIFS.
AI Modeling’s Rapid Progress
The pace at which AI models are improving is part of what makes the technology so exciting for meteorologists. For example, work on the AIFS began in 2022, and in less than four years it has come to outperform other forecasting approaches by 10%–15% on a range of metrics.
According to Irish data scientist Professor Andrew Parnell, this is the equivalent of a decade of research in a couple of years, with the AIFS having since been put into full operational mode just over a year ago.
The computational efficiency of the AI models has also impressed scientists. The speed with which they produce forecasts is "remarkable", Professor Parnell told Irish lawmakers earlier this year, describing them as "a thousand times more efficient" than physics-based approaches.
This efficiency means the models consume less energy and can be more cost-effective than the huge supercomputers required for physics-based modeling, although critics note that such comparisons do not account for the time needed to train ML systems.
Regardless, the bottom line is that AI models are increasingly matching, and in some cases outperforming, their traditional counterparts, while consuming less energy and proving more cost-effective to run.
It is no wonder that scientists are starting to get excited. The Met Office has described AI as the "next frontier" in forecasting, while Professor Parnell has said it has "completely upended" models of forecasting that have been in place for decades.
How Forecasting Has Worked Until Now
Exactly how AI will be integrated into forecasting systems remains a matter of debate. To understand why, it is important to know a little more about how forecasts are currently generated.
Weather forecasting is inherently difficult because it includes so many variables. Consider the summer of 2026, for instance, when the warming of water in a particular region of the Pacific Ocean, a phenomenon known as El Niño, has been linked to elevated temperatures during heatwaves in Europe.
To manage all these variables, meteorologists have for decades relied on physics-based modeling. These models run on large supercomputers that simulate the Earth's system based on its physical laws, performing massive calculations to project how atmospheric conditions will evolve from those currently observed.
As a result of refinements to this form of numerical weather prediction, forecast accuracy has improved by approximately one day per decade. These models have proven dependable and successful.
The Strengths And Limits of AI Models
Although some forms of AI have been incorporated into these models for years, recent advances in areas such as deep learning have paved the way for the creation of independent AI models.
These are trained by being fed very large quantities of historical data, drawn from both real-world observations and physics-based simulations. From this data, they identify inherent patterns that are not explicitly represented, and use them to generate forecasts based on current atmospheric conditions.
This has led to great headline accuracy, meaning the ability to recognize overall patterns, as well as some instances of improved ability to identify unusual or extreme events like the winter storm in Texas in 2021.
The challenge for this type of modeling, however, is that it tends to struggle with local detail, causing features such as a cold front or a storm center to become blurred.
As the AI model strives for a safe prediction by identifying a clear pattern, it tends to smooth out precise features, which reduces forecast usefulness while concealing errors that accumulate over time..
Machine learning models also tend to struggle with extreme weather events, as these, by their very nature, are underrepresented in the available data.
Weighing Risk Against Reward
These fallibilities are precisely why some meteorologists are wary of letting AI models run independently, until the errors can be better understood and accounted for when generating forecasts, or eliminated altogether.
At this early stage in AI modeling, two of the most common proposals for the future use of machine learning have been either to integrate AI into pre-existing physics-based models, or to run them in tandem, allowing their contrasting methods to work in parallel toward more accurate forecasts.
Whichever path meteorological agencies ultimately take, there is broad agreement that AI offers exciting opportunities to improve the accuracy in weather forecasts, and, by extension, to ease the stress for those who depend on such forecasts in their daily lives.
