AI predicted Cyclone Alfred's path and could be the future of forecasting
By meteorologist Tom Saunders and Tyne Logan
Topic:Weather
27m ago27 minutes agoWed 2 Apr 2025 at 7:21pm
Map showing the path of a cyclone heading towards the Queensland coastline.
An AI forecast about Tropical Cyclone Alfred's path was far more accurate than leading weather models. (ABC News)
A week before Tropical Cyclone Alfred was nearing the east coast of Australia, most forecasts were favouring a path either well offshore or near the central Queensland coast.
There was a curious anomaly though: an AI prediction from Google's DeepMind, called Graphcast, was predicting the centre of Alfred would be just 200 kilometres off the coast of Brisbane.
That forecast, made 12 days before ex-Tropical Cyclone Alfred crossed the south-east Queensland coast, was far more accurate than leading weather models used by meteorological organisations around the world, including our own Bureau of Meteorology (BOM).
Two maps showing the trajectory of Cyclone Alfred towards the Queensland coast.
A comparison between the highly regarded ECMWF physical model and AI generated Graphcast. (Supplied: Tristan Meyers)
AI was also outperforming standard models on Alfred's final approach when the system started stalling.
So how can an AI forecast be more accurate than our current weather models, which run on some of the fastest supercomputers on the planet?
AI forecasting vs traditional modelling
A conventional weather model starts with tens of millions of daily observations from around the globe, including from weather balloons, ships, aircraft, ground stations and satellites.
This data is then fed into a model that breaks the globe into a grid both vertically and horizontally, with billions of boxes representing each region of the atmosphere.
A supercomputer then uses the laws of physics, chemistry and fluid dynamics to calculate how each grid box will change into the future.
But AI tackles the problem of weather forecasting in a completely different way — it has no knowledge of science and how the atmosphere works but rather has been trained to find patterns of how the weather behaves based on decades of data.
"And it turns out if you do that over 60 years of data, you learn something that looks approximately like the weather and approximately like the laws of physics, but it was never hard-coded in," University of Oklahoma professor Amy McGovern explained.
This method has some advantages.
Firstly, it's fast and cheap.
While a traditional model is spending hours on a supercomputer crunching trillions of equations, an AI model within minutes on a standard computer can analyse the current weather pattern and use its training to forecast the weather.
So while training the model is expensive, after that it's much more efficient.
"They are very fast. So they are about 1,000 times faster and use less computing resources as the traditional models," director of European Centre for Medium-Range Weather Forecasts (ECMWF) Florence Rabier said.
A middle-aged woman sitting in an office smiling at a camera while wearing a powder blue blazer and black shirt with a necklace.
Florence Rabier says the ECMWF's AI model is especially accurate forecasting the path of tropical cyclones. (Supplied: Dr Florence Rabier)
Are AI models more accurate?
While speed and cost are advantageous, accuracy is the ultimate forecasting goal.
As we saw during Cyclone Alfred, this simplistic machine learning approach is challenging traditional models for precision.
One example though, doesn't demonstrate superiority. However, verification data suggests for some forecasting aspects, AI models are already on top.
Below is a graph comparing the accuracy of weather models. It shows how close the forecasts were in 2022 to what actually played out.
What's clear is the AI temperature forecast has a smaller error than the traditional physical model. In other words, AI's predictions were closer to the observations.
This matches what the ECMWF has found with its AI model, too.
"It's true that AI models tend to be a bit better, sometimes up to 20 per cent better in some parameters." Dr Rabier said.
And an AI model picking Cyclone Alfred appears to be no fluke.
Dr Rabier said the ECMWF's AI model was especially accurate at forecasting the track of tropical cyclones.
"So they are 25 per cent better in that respect … definitely for the track."
But before we suddenly all switch to AI to decide what to wear each day or when to prepare for a cyclone, there's a major caveat — since AI models are trained on past weather, they struggle to forecast unprecedented extreme events, which is critical for protecting lives and property.
AI models also have a disadvantage in their resolution.
For example, the ECMWF's own AI model splits the globe into grids with dimensions three times larger than their physical model — 28km scale vs 9km scale.
That's a decisive hindrance in forecasting both the intensity of weather systems and being able to give a hyperlocal forecast.
A flood in a rural area of Queensland.
AI models may struggle to forecast the intensity of events, like the recent record rain and flooding in Queensland. (Supplied: Roger Volz)
Essentially, AI scores well when the weather is averaged out but won't be able to pinpoint extreme events in localised regions yet, which are often the most dangerous.
"My fear is that people are selling it and convincing people that it works great when it isn't necessarily well tested on every event,"
Professor McGovern said.
Accurate forecasts for the future
If AI has already caught up, what does the future hold?
Professor McGovern believes at the rate AI is improving, it could bring about some really big and exciting leaps in forecasting ability over the next five to 20 years.
Having three-week forecasts with the same accuracy as our current seven-to-10-day forecasts, for example, doesn't seem too far-fetched anymore.
A woman wearing glasses, a powder blue blazer with a white shirt standing in front of a piece of art resembling Oklahoma.
Amy McGovern believes AI could bring some really big and exciting leaps in forecasting ability in the future. (Supplied: Amy McGovern)
"I'm not going to be able to tell you 28 days in advance that there's going to be hail over your street. The best we can do on that is probably an hour or two," Professor McGovern said.
"But I could tell you 28 days in advance maybe that there's a larger-scale event, an extra tropical cyclone or tropical cyclone coming."
So will the BOM be switching over to AI for your daily forecast anytime soon?
According to a BOM spokesperson, they are researching AI but, for now, still trust the old weather models to keep Australians safe.
"Dynamical models, which are based on a deep understanding of the physics of the oceans and atmosphere, provide the foundation of the bureau's forecast capability," they said.
And while AI clearly has a bright outlook, that doesn't mean replacing meteorologists with machines, as there are still many aspects of forecasting, like the impact of weather, that humans are better at.
"In no way do I think meteorologists are going to be out of a job," Professor McGovern said.
Posted27m ago27 minutes agoWed 2 Apr 2025 at 7:21pm
Copy link
X (formerly Twitter)