Kerry Plowright had his feet up and was watching TV one evening late last year when his phone warned of incoming hail.
“I was stunned when I walked out the door because there was just this roar,” he says, describing the sound of hailstones hitting roofs in the New South Wales town of Kingscliff. He had just enough time to move his cars under canvas sails, sparing them from damage.
Plowright isn’t alone in having little warning before wild weather during Australia’s seemingly relentless summer of extremes. This season may include a second tropical cyclone to strike Queensland.
The Albanese government has launched an inquiry into warnings issued by the Bureau of Meteorology and emergency authorities after complaints by councils and others that some alerts lacked accuracy and timeliness.
But Plowright’s case is a little different – his hail heads-up was triggered by data generated by his own firm, Early Warning Network.
Early Warning Network analyses data from radars and remote sensors to detect and issue alerts on extreme heat, rainfall and flooding. It counts local councils and big insurers among its customers.
Private businesses have long offered services based on data from BoM or agencies such as the European Centre for Medium-Range Weather Forecasts (ECMWF). But Early Warning Network is starting to test artificial intelligence models that promise to make a lot more weather information available both rapidly and at low cost.
“You have to pay a bucket load for [ECMWF] data,” Plowright says. “We don’t need now a supercomputer to go and run a forecast that will be extremely accurate up to 10 days, especially for extreme weather.”
Artificial intelligence “is going to be absolutely phenomenal with weather and ultimately climate too, once it gets there”, he predicts.
How AI can help us prepare for weather extremes
Juliette Murphy, a water resources engineer, is similarly excited. 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 traditional physics-based hydrology and hydraulic models. Even relatively basic computers can comb through “really large datasets” quickly to identify likely effects of a flood, she says.
Her clients include Queensland’s fire and emergency services. Its results complement BoM’s, helping authorities decide which homes to evacuate and which roads to close. “That’s important not least because almost half of flood deaths involve people in cars,” Murphy says.
A BoM spokesperson says the bureau had been “proactively and safely engaging with artificial intelligence capabilities for several years”.
“This area of research is one of many initiatives the bureau actively pursues to improve its services to government, emergency management partners and the community,” she says.
Justin Freeman, a computer scientist, ran BoM’s research team which was working on machine learning before he left in in late 2022 to set up his own firm, Flowershift.
Flowershift is building a geospatial model trained on existing observational data. “We would be filling in gaps around what the current forecast products are”, such as providing forecasts in remote regions of Australia or beyond, Freeman says.
“There’s a lot more flexibility to be able to explore things [outside BoM] and use technologies which are very new,” says Freeman, who still does contract work for the bureau. “We’ve got this whole new different class of models which are completely different to what [the bureau had] been running for the last 50 years.”
There are many potential uses for models that can analyse data cheaply and then supply localised information. Farmers, for instance, could ask, “Should I spray my crops this week?” and be told why or why not, Freeman says.
“It hasn’t been that long that we’ve had access to something like ChatGPT,” he says. “Look forward like another two years, five years – it’s just going to accelerate and get better and better.”
The limitations of AI
Some BoM and climate researchers, though, caution how much AI-based models, such Google’s GraphCast or Nvidia’s FourCastNet, can improve on numerical models that churn out a range of probabilities.
“For ‘simple’ weather forecasting and for downscaling physical model data I reckon [there’s] massive potential,” one bureau scientist says. “For warning us of real dangers when the atmosphere gets violent, I’d be very cautious.
“And with climate change, we need to better understand things that are well outside the norm.”
Sanaa Hobeichi, a post-doctoral researcher at the ARC Centre of Excellence for Climate Extremes, says there are still benefits despite the limitations.
Existing climate models typically offer only “coarse” resolutions, such as estimating rainfall changes over areas 150km by 150km. In Sydney, for instance, a model that size would capture the city, mountains and a lot besides and so be of limited use.
Google’s GraphCast forecast model has a resolution down to 28km by 28km, while Hobeichi says some AI can model just 5km by 5km.
A challenge, though, is that machine-learning techniques inherit and potentially extrapolate imperfections of the traditional models they train on.
Jyoteeshkumar Reddy Papari, a post-doctoral CSIRO researcher, notes that the ECMWF was initially sceptical of AI but has lately started its own experimental model. It is also displaying several others on its website, including Google’s.
“Countries that don’t have good meteorological organisations are relying on these machine learning models because they are super easy to learn and are publicly available,” he says. “So some of the African countries are using these forecasts.”
Google researchers last year claimed GraphCast “significantly outperforms the most accurate” operational systems in 90% of 1380 targets. Tropical cyclones, atmospheric rivers and extreme temperatures were predictions it made which were better than traditional models and improvements are ongoing.
“One particular example we often mention was Hurricane Lee, because it was the first time that we observed in real time how GraphCast was predicting a hurricane trajectory that originally differed from the traditional systems, and eventually was shown to be the right trajectory,” said Alvaro Sanchez-Gonzalez, a Goggle researcher. “It was detected in real time and it was verified by independent sources.”
Current tracking 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.
ECMWF’s machine learning coordinator, Matthew Chantry, says AI models are “a very exciting avenue as a companion system for traditional forecasting” although the latter retains some advantages.
“Tropical cyclone intensity estimates are a good example,” he says. “It’s an open question whether these flaws are maintained as the technology matures – it is still very early days.”
Authorities act based on the probabilities calculated by traditional models but that needs a very large supercomputer. “With AI forecasts, this is dramatically reduced, with some estimates suggesting a 1000-times reduction in the energy to make a forecast. Cheaper systems could therefore be a force for equality.
“This reduced cost could also be invested into larger ensembles, meaning that we have a better idea of low-probability but extreme events that could occur.”
And as for predicting effects of a heating planet?
“The problem is significantly harder than weather forecasting, with less data,” says Chantry. “That said, in a changing climate, where evidence suggests an increase in extreme events, then any help with predicting these events has significant value.”
