Kerry Plowright was watching TV one evening late last year when his phone warned of approaching hail.
“I was stunned when I walked out the door because there was just a roar,” he says, describing the sound of hailstones hitting rooftops in the New South Wales town of Kingscliff. He had just enough time to move his cars under canvas tarps so they weren’t damaged.
Plowright isn’t the only one with little warning of wild weather during Australia’s seemingly brutal summer of extremes. A second tropical cyclone could hit Queensland this season.
Related: Cyclone forecasts, boosted by artificial intelligence, allow the path to be tracked sooner
The Albanian government has launched an investigation into warnings from the Bureau of Meteorology and emergency authorities after complaints from municipalities and others that some warnings were not accurate and timely.
But Plowright’s case is a little different: his heads-up was caused by data generated by his own company, Early Warning Network.
Early Warning Network analyzes data from radars and external sensors to detect and issue warnings about extreme heat, rainfall and flooding. It counts municipalities and large insurers among its customers.
Private companies have long offered services based on data from the BoM or agencies such as the European Center for Medium-Range Weather Forecasts (ECMWF). But Early Warning Network is starting to test artificial intelligence models that promise to make much more weather information available quickly and at low cost.
“You have to pay a bucket load for it [ECMWF] data,” says Plowright. “We now don’t need a supercomputer to make predictions that are extremely accurate for up to ten days, especially in extreme weather.”
Artificial intelligence “will be absolutely phenomenal in terms of weather and ultimately climate, once it arrives,” he predicts.
How AI can help us prepare for extreme weather events
Juliette Murphy, a water resources engineer, is also enthusiastic. She founded FloodMapp to give communities more time to prepare after monitoring devastating floods in Queensland’s Lockyer region in 2011 and then in the Canadian city of Calgary two years later.
FloodMapp uses machines that learn from each model run, as well as from traditional physics-based hydrology and hydraulic models. Even relatively simple computers can quickly sift through “really large data sets” to identify the likely impacts of a flood, she says.
Her clients include Queensland’s fire and emergency services. The results complement those of the BoM, helping authorities decide which houses to evacuate and which roads to close. “That’s important not least because almost half of flood deaths involve people in cars,” says Murphy.
Related: Flood warnings prompt evacuations in NT as Queensland braces for an approaching cyclone
A spokesperson for the BoM said the agency “has been proactively and securely working on artificial intelligence for several years”.
“This area of research is one of many initiatives the agency is actively pursuing to improve its services to government, emergency management partners and the community,” she said.
Justin Freeman, a computer scientist, led BoM’s research team working on machine learning before leaving at the end of 2022 to set up his own company Flowershift.
Flowershift builds a geospatial model trained on existing observational data. “We would fill gaps around what the current forecast products are,” such as providing forecasts in remote areas of Australia or beyond, Freeman says.
“There’s a lot more flexibility to be able to explore things [outside BoM] and using technologies that are very new,” says Freeman, who still does contract work for the agency. “We have a whole new class of models that are completely different than anything else [the bureau had] has been active for fifty years.”
There are many potential applications for models that can analyze data cheaply and then provide localized information. For example, farmers might ask themselves, “Should I water my crops this week?” and you are told why or why not, says Freeman.
“It wasn’t that long ago that we had access to something like ChatGPT,” he says. “Look ahead two more years, five years – it’s just going to accelerate and get better and better.”
The limitations of AI
However, some BoM and climate researchers caution how much AI-based models, such as Google’s GraphCast or Nvidia’s FourCastNet, can improve numerical models that produce a range of probabilities.
Related: ‘Very scary’: Mark Zuckerberg’s promise to build advanced AI alarm experts
“For ‘simple’ weather forecasting and for reducing physical model data, I think [there’s] huge potential,” says an agency scientist. “If I were to warn us of real dangers if the atmosphere turns violent, I would be very careful.”
“And with climate change, we need to better understand things that are far outside the norm.”
Sanaa Hobeichi, a postdoctoral researcher at the ARC Center of Excellence for Climate Extremes, says there are still benefits despite the limitations.
Existing climate models generally only provide ‘coarse’ resolutions, such as estimating changes in precipitation over areas of 150 by 150 km. In Sydney, for example, a model of that size would capture the city, mountains and more and would therefore be of limited use.
Google’s GraphCast prediction model has a resolution of just 28 by 28 km, while Hobeichi says some AI can only model 5 by 5 km.
One challenge, however, is that machine learning techniques inherit and potentially extrapolate imperfections from the traditional models they train on.
Jyoteeshkumar Reddy Papari, a CSIRO postdoctoral researcher, notes that the ECMWF was initially skeptical of AI, but recently started its own experimental model. It also shows several others on its website, including Google’s.
“Countries that don’t have good meteorological organizations rely on these machine learning models because they are super easy to learn and publicly available,” he says. “So some African countries are using these predictions.”
Google researchers claimed last year that GraphCast “significantly outperforms the most accurate” operational systems in 90% of 1,380 targets. Tropical cyclones, atmospheric rivers and extreme temperatures were forecasts that beat traditional models and improvements are still underway.
Related: AI will impact 40% of jobs and likely worsen inequality, says IMF head
“A specific example we often mention is Hurricane Lee, because it was the first time we observed in real time how GraphCast predicted a hurricane track that was originally different from traditional systems, and ultimately turned out to be the correct track,” says Alvaro Sanchez-Gonzalez, a Goggle researcher.
“It was detected in real time and it was verified by independent sources.”
Current tracking of the potential Coral Sea cyclone – which will be named Kirrily if it forms as expected on Monday – will also be monitored to see how models compare.
Matthew Chantry, ECMWF’s machine learning coordinator, says AI models are “a very exciting way to act as a companion system to traditional forecasting,” although the latter retains some advantages.
“Tropical cyclone intensity estimates are a good example,” he says. “It is an open question whether these shortcomings will persist as the technology matures – it is still in its infancy.”
Authorities act based on the probabilities calculated by traditional models, but this requires a very large supercomputer. “With AI predictions, this is dramatically reduced, with some estimates suggesting a thousand-fold reduction in the energy to make a prediction. Cheaper systems can therefore be a driving force for equality.
“These lower costs could also be invested in larger ensembles, meaning we have a better idea of low-probability but extreme events that could occur.”
And as for predicting the effects of a warming planet?
“The problem is significantly more difficult than weather forecasting, with less data,” says Chantry. “That said, in a changing climate, where evidence points to an increase in extreme events, any help in predicting these events would be of great value.”