AI Could Set a New Bar for Designing Hurricane-Resistant Buildings

NIST postdoctoral researcher Rikhi Bose, together with Pintar and NIST Fellow Emil Simiu, used these new techniques and resources to tackle the issue from a different angle. Rather than having their model mathematically build a storm from the ground up, the authors of the new study taught it to mimic actual hurricane data with machine learning, Pintar said. 

Studying for a physics exam by only looking at the questions and answers of previous assignments may not play out in a student’s favor, but for powerful AI-based techniques, this type of approach could be worthwhile. 

With enough quality information to study, machine-learning algorithms can construct models based on patterns they uncover within datasets that other methods may miss. Those models can then simulate specific behaviors, such as the wind strength and movement of a hurricane.

In the new research, the study material came in the form of the National Hurricane Center’s Atlantic Hurricane Database (HURDAT2), which contains information about hurricanes going back more than 100 years, such as the coordinates of their paths and windspeeds.

The researchers split data on more than 1,500 storms into sets for training and testing their model. When challenged with concurrently simulating the trajectory and wind of historical storms it had not seen before, the model scored highly. 

“It performs very well. Depending on where you’re looking at along the coast, it would be quite difficult to identify a simulated hurricane from a real one, honestly,” Pintar said. 

They also used the model to generate sets of 100 years’ worth of hypothetical storms. It produced the simulations in a matter of seconds, and the authors saw a large degree of overlap with the general behavior of the HURDAT2 storms, suggesting that their model could rapidly produce collections of realistic storms. 

However, there were some discrepancies, such as in the Northeastern coastal states. In these regions, HURDAT2 data was sparse, and thus, the model generated less realistic storms. 

“Hurricanes are not as frequent in, say, Boston as in Miami, for example. The less data you have, the larger the uncertainty of your predictions,” Simiu said.

As a next step, the team plans to use simulated hurricanes to develop coastal maps of extreme wind speeds as well as quantify uncertainty in those estimated speeds.  

Since the model’s understanding of storms is limited to historical data for now, it cannot simulate the effects that climate change will have on storms of the future. The traditional approach of simulating storms from the ground up is better suited to that task. However, in the short term, the authors are confident that wind maps based on their model — which is less reliant on elusive physical parameters than other models are — would better reflect reality. 

Within the next several years they aim to produce and propose new maps for inclusion in building standards and codes.