Detailed predictions about how an approaching tsunami will impact Japan’s northeastern coastline can now be made in fractions of a second. This potentially life-saving technology makes use of the capabilities of machine learning.
On March 11, 2011, a devastating tsunami struck northeast Japan. It killed approximately 18,500 people. Many lives could have been saved if early warning of the approaching tsunami have been given. Accurate predictions of how high the water would rise at various points along the coastline and further inland would have saved many lives.
The coast now has the world’s largest network of sensors for monitoring ocean floor movement. This network’s 150 offshore stations provide tsunami warnings in advance. However, for the data generated by the sensors to be meaningful, it must be converted into tsunami heights and extents along the coastline.
This usually necessitates solving difficult nonlinear equations numerically. It takes about 30 minutes on a standard computer. However, the 2011 tsunami struck some parts of the coast only 45 minutes after the earthquake.
Iyan Mulia of the RIKEN Prediction Science Laboratory and colleagues have now reduced the calculation time to less than one second using machine learning.
“The main benefit of our method is the speed of prediction. Because it is critical for early warning,” Mulia explains. “Traditional tsunami modelling predicts after 30 minutes, which is too late. However, our model can make predictions in seconds.”
Because tsunamis are uncommon, the team used over 3,000 computer-generated tsunami events to train their machine-learning system. They then put it through its paces with 480 other tsunami scenarios and three actual tsunamis. Their machine-learning-based model achieved comparable accuracy while requiring only 1% of the computational effort.
The same deep-learning approach could be used in other disaster scenarios where time is critical. “The sky is the limit. You can apply this method to any kind of disaster prediction where the time constraint is very short,” says Mulia. Mulia became interested in tsunami research after the 2004 Indian Ocean tsunami devastated coastal areas in his home country of Indonesia.
The findings have been published in the journal Nature Communications.
Mulia points out that the method is only accurate for large tsunamis of more than 1.5 metres in height. So, he and his colleagues are now working to improve its accuracy for smaller tsunamis.
More information: Iyan E. Mulia et al, Machine learning-based tsunami inundation prediction derived from offshore observations, Nature Communications (2022). DOI: 10.1038/s41467-022-33253-5
