Tech giants are rolling out a wave of machine learning weather forecasting models, challenging traditional physics-based models developed over decades. But are these models actually accurate?
Article contents:
- Introduction
- Traditional methods
- AI revolution
- Performance comparison
- Limitations and challenges
- Competitive advantages
- Future outlook
- Conclusion
Introduction
A wave of machine learning AI weather forecasting models has been launched by some of the largest companies on the planet. These are challenging the orthodoxy of traditional physics-based computer forecasting, which has been developed and incrementally improved over many decades. But are machine learning models any good?
Revolutionary discovery: Aardvark Weather predicts the weather with minimal resources
The weather is a national obsession for us Brits, and it’s little wonder given the dramatic changes seen and felt from day to day. Accurate weather forecasts are not only vital for planning our daily lives, but knowing severe weather is coming can help us change our behavior, save lives, and mitigate damage to property.
Traditional methods
It is impossible to evaluate the full economic value of weather forecasting globally, but the figures are massive. According to the National Oceanic and Atmospheric Administration, in the United States alone—and considering only the largest weather disasters that caused over $1 billion in damages—the fallout of severe weather in 2024 totaled $182 billion, with 568 fatalities.
Since 1980, that damage figure stands at roughly $3 trillion! Meanwhile, in the UK, there were 1,311 excess deaths caused by heatwaves in 2024. A study by London Economics consultants concluded that the Met Office will bring £56 billion of benefits to the British economy over a decade through the provision of meteorological services.
Traditional weather forecasts are produced on some of the largest supercomputers on the planet; the Met Office’s supercomputing contract is worth £1.2 billion. This vast sum of money buys you a machine capable of performing 60 quadrillion (60,000,000,000,000,000) calculations per second.
AI revolution
Machine learning weather forecasting models have only been around for a few years, and they are showing promise while developing rapidly. Traditional models take hours to run on expensive supercomputers, yet this new generation of models can take less than a minute to run on an ordinary laptop. They do not need to know all the cumbersome laws of physics, but are instead trained on 40 years of past data to make their forecasts.
How do they perform? Well, let’s look at forecast verification data from the European Centre for Medium-Range Weather Forecasts for atmospheric pressure patterns in the 2024/2025 winter. GraphCast (Google), AIFS (ECMWF), and Aurora (Microsoft) were more accurate than the traditional standard IFS (ECMWF) forecast, while FourCastNet (Nvidia) and Pangu-Weather (Huawei) lagged behind.
Performance comparison
So some machine learning models performed better, and some worse, but it depends on which variable you look at, and all of this could change quickly as the rate of progress accelerates. Just like traditional models, AI models are less accurate the further into the future you try to forecast—a result of the chaotic nature of the atmosphere.
Looking 10 days ahead, none of the AI models (or traditional models) were able to provide forecasts considered to be of much use in terms of accuracy. So is it time to abandon physics-based atmospheric models? Not yet!
Limitations and challenges
AI weather forecasting machine learning models are not only trained using data produced by traditional atmospheric models, but they also use the initial state of the atmosphere from traditional models as their input point. In other words, without running those traditional models, machine learning models will not work as well.
All machine learning models can predict large-scale features like high and low pressure six days ahead very well, but their performance can drop compared to traditional models at scales smaller than 1,000 kilometers or less. This means important features like troughs and ridges can be missed, making the difference between a dry day or a day with heavy rain.
Competitive advantages
The majority of machine learning models have a resolution of 28 square kilometers, which is the same scale as the data they were trained on. This means small features like showers will likely be missed, so they won’t be able to predict a Boscastle flood event many days in advance. News headlines have claimed these new models are better than traditional models at predicting hurricanes.
It may be true that some were slightly better at predicting hurricane landfalls ahead of traditional models, but at the same time, they were very poor at predicting wind strength and therefore the potential damage the storm would bring. This may be the result of a smoothing effect, or averaging from looking at too many hurricanes in 40 years of training data.
Future outlook
AI models may struggle to predict the impacts of rare events not seen very often in 40 years of training data. An example of this is the 1991 Mount Pinatubo eruption, which cooled the planet by up to 0.5 degrees Celsius for two years. There are also questions about how well AI models will perform in forecasting a warmer world as our planet continues to warm due to climate change.
Past climate they were trained on will look very different from our future climate, as greenhouse gases continue to build up in the atmosphere. Professor Kristin Dale, chief AI officer at the Met Office, says: “I think we will have traditional models working hand-in-hand with AI models so that we leverage their combined strengths to enable hyper-local accurate forecasts, delivered quickly, when you need them.”
Conclusion
Machine learning models haven’t been around for very long, but with their speed, computer efficiency, and rapid development rate, they are showing great potential. The future looks promising for combining traditional models with AI models to get the best of both worlds. As technology continues to evolve, we may witness a true revolution in the accuracy and speed of AI weather forecasting, benefiting society as a whole and helping protect lives and property from severe weather phenomena.
