Two technologies, one powerful combo
Hoping to prevent a replay of the Great East Japan Earthquake, Fujitsu, Tohoku University and the University of Tokyo joined forces to harness artificial intelligence (AI) and high performance computing to take tsunami forecasting to the next level—with their work representing the culmination of a decade’s worth of research efforts.
For instance, in a 2015 Geophysical Research Letters publication, the Oishi-led team described the development of a parallel version of the TUNAMI-N2 (Tohoku University’s Numerical Analysis Model for Investigation of Near-field Tsunamis) model that could be run on 9,469 cores of the K computer. Their five-meter resolution simulation of flooding conditions in Sendai, the largest city in the Tohoku region, on the supercomputer took only 93.2 seconds; the same computations would have taken several days on a workstation.
Applying their findings to the 2011 disaster, it took the model five minutes to analyze the tsunami wave source and ten minutes to provide basic flood predictions. As the tsunami took an hour to reach Sendai, having such information publicly available so shortly after the quake could have turned the tide in terms of survival.
Despite this breakthrough in tsunami forecasting, the model’s computational demands make it largely inaccessible and impractical should such a scenario happen again. To this end, Oishi and colleagues later proposed leveraging a convolutional neural network (CNN) for end-to-end tsunami flooding forecasting, publishing their findings in Nature Communications in early 2021.
To develop the CNN, a form of artificial neural network used in image recognition and processing, the team conducted tsunami simulations for a total of 12,000 scenarios—synthetically generating 10,000 cases for training and then evenly splitting the remaining 2,000 cases between validation and testing. The network’s training was performed in the AI Bridging Cloud Infrastructure, a GPU-accelerated, 226-petaFLOPS supercomputer operated by the National Institute of Advanced Industrial Science and Technology in Japan.
Incredibly, the trained CNN took only 0.004 seconds on average to provide a forecast using a single CPU node with 40 cores. Compared to the previous 1.5-minute standard set by the K computer, the team’s approach was not only much faster, but also required fewer computational resources.
“AI-based tsunami forecasting makes it possible to predict the state of flooding directly from observed offshore tsunami waveforms, without relying on tsunami source estimation results,” noted Oishi. “While the AI requires large amounts of calculation during training, in the event of a disaster, it can predict tsunami flooding
immediately.”
As impressive as their achievements in tsunami forecasting may sound, the team shows no sign of stopping.
“By further utilizing the large-scale, high-speed performance of Fugaku and training the system with additional tsunami scenarios, Fujitsu aims to realize an AI that can offer predictions for unexpected tsunami and flooding predictions over a wider area,” added Oishi.
