Machine Learning Algorithms for Real-time Tsunami Inundation Forecasting: A Case Study in Nankai Region

Abstract

The most acceptable method to estimate tsunami inundation caused by a submarine earthquake is by conducting a nonlinear tsunami simulation. However, this method has the disadvantages of a relatively high computational cost and the necessity for immediate warning announcements when a tsunami is imminent. In this study, to overcome this problem, we apply two machine learning models, a convolutional neural network and a multilayer perceptron, to estimate tsunami inundation in real time. We run multiple fault scenarios and store the result of the maximum tsunami amplitude in a low-resolution grid and the associated tsunami inundation in a high-resolution grid in the database. The convolutional neural network selects tsunami inundation in the high-resolution grid as the forecast based on pattern similarity between the input, which is the results of linear forward modeling in the low-resolution grid, and the precomputed patterns in the database. Slightly different from the convolutional neural network, instead of selecting the best-fit scenario in the database, the multilayer perceptron directly generates the inundation forecast based on knowledge acquired during the training process. We conduct an experiment using the hypothetical future Nankai megathrust earthquake with Atashika and Owase Bays in Japan as the study cases. The results show that our proposed methods are extremely fast (less than 1 s) and comparable with nonlinear forward modeling. Therefore, the proposed methods can be used as a deterministic model for real-time simulation.