Turning The Tide With AI And HPC

By harnessing both artificial intelligence and high performance computing in one powerful model, scientists from Japan are making real-time tsunami prediction more accessible.



The future of forecasting

True enough, by February 2021, the same team announced that they had successfully leveraged Fugaku to streamline the development of a new AI that can instantaneously predict tsunami flooding even on regular computers. Building upon their earlier study, the researchers used the supercomputer to generate 20,000 possible tsunami scenarios based on high-resolution, three-meter unit simulations.

They then trained the AI on these datasets, allowing the team to build a model that first roughly approximates land flooding based on offshore observations of tsunami waveforms. The algorithm then increases the resolution of estimated flooding conditions and optimizes calculations, enabling the prediction of floods in coastal areas before landfall at high spatial resolutions.

To put their new technology to the test, Oishi and colleagues sought to glimpse into potential tsunami flooding in Tokyo Bay caused by a theoretical quake in the Nankai Trough. Earthquakes of magnitude 8.0 and higher have occurred repeatedly in the trough every 100 to 200 years, with the next one anticipated by the Japanese government to likely happen within the upcoming 30 years.

“A magnitude 9.0 earthquake can be expected in the case of massive earthquake in the Nankai Trough,” he said. “Therefore, we built an AI model based on training data with such a magnitude.”

Running the simulation on just a desktop computer, the AI generated highly accurate tsunami flood forecasts for the coastal urban areas of Tokyo Bay in a matter of seconds. Their results were comparable to the tsunami model used by the Cabinet Office of Japan, which is known among seismologists as the largest possible earthquake model built, according to Oishi.

“By generating training data of 20,000 cases using Fugaku, we also predicted tsunamis three times the height of the Cabinet Office’s model,” he added.

As the model can be run on conventional computers, practical, real-time flood prediction systems are now within reach even in locales that lack access to powerful supercomputers not just in Japan, but also potentially in other disaster-prone countries. Such an approach will hopefully allow disaster management teams to make better data-driven mitigation and evacuation measures in the future—minimizing the loss of life and preventing a replay of the Great East Japan Earthquake.

“In AI-based tsunami prediction, properly designed training data is important. In the future, we will further leverage Fugaku’s high performance and high speed to generate various tsunami scenarios with the aim of predicting ‘unexpected’ tsunamis,” concluded Oishi.

 

This article was first published in the print version of Supercomputing Asia, January 2022.
Click here to subscribe to Asian Scientist Magazine in print.

———

Copyright: Asian Scientist Magazine. Illustration: Shelly Liew/Asian Scientist Magazine.
Disclaimer: This article does not necessarily reflect the views of AsianScientist or its staff.

A molecular biologist by training, Kami Navarro left the sterile walls of the laboratory to pursue a Master of Science Communication from the Australian National University. Kami is the former science editor at Asian Scientist Magazine.

Related Stories from Asian Scientist