ALBERT

Description: Machine learning is one of the methods employed for the immediate prediction of tsunami occurrence. There have been few studies on error evaluation focusing on the arrival time of the tsunami inundation. In this study, we used machine learning to predict the tsunami arrival times for two areas and evaluated the errors. Tsunamis were calculated for several thousand cases for three different earthquake sizes using the tsunami simulator Q-Wave and assuming a Nankai Trough earthquake. Using these data, we trained a neural network to predict tsunami arrival time on the basis of the initial water level. As the amount of training data increased, the error decreased and the model with the most training data was used to predict the tsunami arrival times for two areas. The errors tended to be particularly large in coastal areas where tsunamis often reach. In some cases, the error was larger in Area B than in Area A, even though the amount of data for Area B was double that of Area A. These errors may be ascribed to variations in the arrival time of tsunamis in the training data due to topographical characteristics. For a magnitude 9.0 earthquake, the standard deviation of the tsunami arrival time in Area B is approximately 1.5 times that of Area A. Therefore, in addition to increasing the number of output variables and the accuracy of tsunami simulation, a more detailed consideration of the effect of area-specific characteristics on the error is needed in the future.