by Kerry Plowright had his feet up and was watching TV one evening late last year when his phone alerted him of hail.
“I was shocked when I walked out the door because it was just this roar,” he says, describing the sound of hail hitting roofs in the New South Wales town of Kingscliff. He had just enough time to move his cars under canvas sails, which prevented them from damage.
Plowright isn’t the only one sounding wild weather warnings during Australia’s seemingly unrelenting summer of extremes. This season may include a second tropical cyclone to hit Queensland.
Related: Artificial intelligence-enhanced cyclone forecasting offers earlier path tracking
The Albanian government has launched an inquiry into warnings issued by the Bureau of Meteorology and emergency authorities following complaints from councils and others that some warnings lacked accuracy and timeliness.
But Plowright’s case is slightly different – his bosses were motivated by data created by his own firm, Early Warning Network.
An Early Warning Network analyzes data from radar and remote sensors to detect and issue alerts for extreme heat, rain and flooding. Its customers include local councils and major insurers.
Private firms have long offered services based on data from the Management Board or from agencies such as the European Center for Medium-Range Weather Forecasts (ECMWF). But the Early Warning Network is starting to test artificial intelligence models that promise to provide much more weather information quickly and at low cost.
“You have to pay a bucket load for it [ECMWF] details,” says Plowright. “We don’t need a supercomputer now to run a forecast that’s incredibly accurate up to 10 days, especially for extreme weather.”
Artificial intelligence is “going to be really great with the weather and eventually the climate, when it’s there”, he predicted.
How AI can help us prepare for extreme weather conditions
Juliette Murphy, a water resources engineer, is similarly excited. She founded FloodMapp to give communities more time to prepare after monitoring devastating floods in the Lockyer region of Queensland in 2011 and then in the Canadian city of Calgary two years later.
FloodMapp uses machines that learn from all models run as well as traditional hydrological and physics-based hydrological models. Even fairly basic computers can quickly comb through “really large data sets” to identify the likely effects of a flood, she says.
Her clients include the Queensland fire and emergency services. Its findings complement the BoMs, helping authorities decide which houses to evacuate and which roads to close. “That’s especially important because nearly half of flood deaths involve people in cars,” says Murphy.
Related: Flood warnings prompt evacuations in the NT as Queensland braces for an incoming cyclone
A BoM spokesperson says the bureau has been “proactively and safely engaging with artificial intelligence capabilities for several years”.
“This area of research is one of many initiatives the bureau is actively pursuing to improve its services to government, emergency management partners and the public,” she says.
Justin Freeman, a computer scientist, ran BoM’s research team working on machine learning before leaving in late 2022 to found his own firm, Flowershift.
Flowershift is building a geospatial model trained on current observational data. “We would be filling in gaps around what the current forecasting products are”, such as providing forecasts in remote regions of Australia or further afield, says Freeman.
“There’s a lot more flexibility to be able to explore things [outside BoM] and uses technologies that are very new,” says Freeman, who still does contract work for the bureau. “We have this whole new different class of models that are completely different to what they are [the bureau had] has been going on for the past 50 years.”
There are many potential uses for models that can freely analyze data and then provide local information. Farmers, for example, might ask, “Should I spray my crops this week?” and to tell you why or wherefore, says Freeman.
“It wasn’t that long ago that we had access to something like ChatGPT,” he says. “Look ahead like another two, five years – it’s going to accelerate and get better and better.”
Limits of AI
Some BoM and climate researchers warn, however, how much AI-based models, such as Google’s GraphCast or Nvidia’s FourCastNet, can improve on numerical models that feed a range of probabilities.
Related: ‘Very scary’: Mark Zuckerberg’s promise to build advanced AI alarm experts
“In the case of ‘simple’ weather forecasting and to lower the data of physical models I think [there’s] huge potential,” says one bureau scientist. “To warn us of real dangers when the atmosphere becomes violent, I would be very careful.
“And with climate change, we need to better understand things that are way outside the norm.”
Sanaa Hobeichi, a postdoctoral researcher at the ARC’s Climate Center of Excellence for Climate Extremes, says there are still benefits despite the limitations.
Current climate models typically only offer “coarse” solutions, such as estimating rainfall changes over 150km by 150km areas. In Sydney, for example, such a large model would include the city, the mountains and much more and therefore would have limited use.
Google’s GraphCast forecast model has resolution down to 28km by 28km, and Hobeichi says some AI can model just 5km by 5km.
A challenge, however, is that machine learning techniques inherit and can extrapolate the flaws of the traditional models they train.
Jyoteeshkumar Reddy Papri, a CSIRO postdoctoral researcher, notes that the ECMWF was initially skeptical of AI but has recently started its own experimental model. Many others are also displayed on his website, including the Google site.
“Countries that don’t have good meteorological organizations are relying on these machine learning models because they are extremely easy to learn and publicly available,” he says. “So some of the African countries are using these forecasts.”
Google researchers claimed last year that GraphCast “significantly outperforms the most accurate operating systems” in 90% of 1380 targets. His predictions of tropical cyclones, atmospheric rivers and extreme temperatures were better than traditional models and improvements are still underway.
Related: AI will affect 40% of jobs and likely to increase inequality, IMF chief says
“One specific example we often cite is Hurricane Lee, because it was the first time we observed in real time how GraphCast was predicting a hurricane’s trajectory that was different from traditional systems at first, and that was shown in the it ends up being the right trajectory,” said Alvaro Sanchez-Gonzalez, Goggle researcher.
“It was detected in real time and it was verified by independent sources.”
The current track of the potential cyclone in the Coral Sea – to be named Kirrily if it forms as expected by Monday – will also be monitored to see how models compare.
Matthew Chantry, ECMWF’s machine learning co-ordinator, says that AI models are “a very exciting alternative to traditional forecasting” although the latter has some advantages.
“Tropical cyclone intensity estimates are a good example,” he says. “Whether these flaws are maintained as the technology matures is an open question – it’s still very early days.”
Authorities act based on probabilities calculated through traditional models but which require a supercomputer. “With AI forecasts, this has been greatly reduced, with some estimates implying a 1000-fold reduction in forecasting energy. Cheaper systems could be an equalizing force.
“This reduced cost could also be invested in larger ensembles, meaning we have a better idea of low-probability but extreme events that may occur.”
And how do you predict the effects of a warming planet?
“The problem is much more difficult than weather forecasting, with less data,” says Chantry. “That said, in a changing climate, where evidence suggests an increase in extreme events, any help in predicting these events is of significant value.”
