Some AI researchers in the Pacific Northwest think so.
For years now, AI has been receiving attention as a potential new way to save people from natural disasters — particularly where disasters overlap with existing technology. As one Stanford PhD student told The New York Times back in 2019, AI might be a tool to help “anticipate, absorb and recover from events that cause grid outages,” including hurricanes. But AI-based weather modeling — some of which is being designed in the Pacific Northwest — might also be able to fine-tune meteorologists’ overall predictive power at those moments when precision is most crucial.
And with each passing year, this work becomes more pressing.
“We are currently using the model to study the changing hurricane risks under climate change,” Wenwei Xu, a data and climate scientist at the Richland, Washington-based Pacific Northwest National Laboratory, a joint venture run by Battelle and the U.S. Department of Energy told pnw.ai. Last year, Xu and a group of his fellow scientists produced a model for predicting increases in hurricane intensity. This may sound a bit unhelpful, like a police scanner that predicts when in-progress shootings will see upticks in deadliness, but that doesn’t tell the whole story.
2019’s Hurricane Dorian, for instance, was a nightmare for forecasters, and not just because President Donald Trump famously predicted that it would clobber Alabama — an out-of-left-field and utterly false prediction that turned out to be based on a random blob he himself had reportedly drawn on a NOAA weather map with a sharpie. That incident overshadowed the storm’s other big surprise: abruptly increasing in intensity right before it caused record-shattering damage in the Bahamas — where unresolved damage from Dorian remains a low-simmering national tragedy to this day.
Jason Dunion, a NOAA meteorologist, told Wired at the time that Dorian’s sudden surge in intensity was a shock. “Usually you don’t see rapid intensification in just one day, usually that happens to weaker storms because they have more room to grow,” he told that publication, noting that a transition from a Category 4 storm to Category 5 over the course of just a few hours, “was something that doesn’t happen very often.”
And after six consecutive years of “above average” storm seasons, with 2022 shaping up to be the seventh in a row, can the kind of intensification we saw with Dorian be accurately predicted with the improved model designed by the Pacific Northwest National Laboratory? There’s no telling yet, mostly because no one has tried. “I have not heard that [our] model is being used by emergency management departments yet,” Xu said.
The model, which is available on GitHub right now, might well be a leap forward in intensification prediction. Past work on the topic includes what the Pacific Northwest National Laboratory paper calls “a few pioneer studies” applying neural networks to cyclone intensity forecasting, but those were in areas outside the hurricane-forming North Atlantic Basin, and the work was performed prior to the advent of graphics processing units (GPUs). “Hence, the trained networks were much smaller in terms of the number of nodes and the number of trainable weights,” the paper notes. Another promising study from 2019 used similar methodology to last year’s paper, but the authors of that paper demonstrated “no significant improvement” over more traditional predictive methods according to Xu and colleagues.
Their system, a feedforward artificial neural network of the type sometimes known as “multilayer perceptrons,” was fed predictor data from the Statistical Hurricane Intensity Prediction Scheme in order to understand the prediction-outcome relationship between all the Atlantic Basin storms from 2010 through 2018. Their experiment simply involved extracting a year of real data from the model, creating data for a bunch of fake storms to fill in the gap, and comparing the fake year to the one that actually happened.
The storm intensity data it predicted ranged from 9-20% more accurate than the predictions of other models.
Karthik Balaguru, another climate and data scientist at the Pacific Northwest National Laboratory who worked on the model, said there’s still room for improvement. Much of the data forecasters use is from the Argo Program and its global network of floating data-gathering pods, he said. As powerful as all those measurements are, “we still don’t have sufficient data, and the initial conditions of forecast models are associated with errors that propagate into forecasts,” he told pnw.ai.
But, Balaguru said, “real-time observations of the ocean from satellites can fill this gap.” In a prior study, he said he and his colleagues demonstrated that “incorporating salinity from satellites can improve the prediction of hurricane rapid intensification.”
Nonetheless, Xu would like someone to give the existing model a try and see if it helps meteorologists do their lifesaving work. As he said, “I do think it is ready to be used for tropical cyclone intensity forecasts in an emergency preparation setting.”