Last year, Riken launched Fugaku, the world’s fastest supercomputer, a year ahead of schedule. The system launched early in order to combat a disaster threatening Japan: COVID-19. Now, with the pandemic appearing to wane – in no small part due to the help of supercomputers like Fugaku – the massive system is being tasked with another disaster that regularly threatens the country: tsunamis.
Almost exactly ten years ago, a 9.0-magnitude earthquake set off a devastating tsunami that struck the coast of Tōhoku. Nearly 16,000 people were killed, thousands more were injured and thousands remain missing – and beyond the human cost, hundreds of billions of dollars in damage made it the most expensive natural disaster in known human history.
Needless to say, it would have been good to have some warning.
Fujitsu – which jointly developed Fugaku with Riken – partnered with the International Research Institute of Disaster Science at Tohoku University and the Earthquake Research Institute at the University of Tokyo for the new initiative. In partnership, the three organizations have developed an AI model to predict flooding from tsunamis in “near real-time.”
The model processes incoming data on offshore tsunami waveforms, combined with high-resolution data on the coastlines, to generate precise, near-instant predictions of how a given tsunami will flood a given area, including impacts on buildings and roads.
Training data for the model was generated on Fugaku, whose 158,976 nodes and 415 Linpack petaflops simulated a series of 20,000 tsunamis, their ensuing waveforms and the resulting flooding on land. These simulations were accelerated by Fugaku’s built-in AI features and optimization.
Using the data from these 20,000 simulations, the researchers then train the AI model to understand the relationship between waveforms and flooding. This work on heavy-duty hardware pays off when the model is actually deployed: the fully trained model can be run on ordinary computers in mere seconds, and its results have been verified using comparisons with the tsunami model used by the Cabinet Office of Japan.
This rapid flooding prediction based on early waveforms can be used for correspondingly rapid evacuation notices and disaster preparation. This will mark a major shift from previous methods, which relied on crude databases of earthquakes and flooding to provide coarse predictions.
The tsunami project was granted access to Fugaku through Riken’s program for “preliminary use projects,” applications for which were solicited in advance of its general availability.