A Neural Network-Based Surrogate Model for Efficient Probabilistic Tsunami Inundation Assessment

Abstract

Probabilistic assessment or uncertainty evaluation of the inundation depth and distribution of tsunamis are critical for effective tsunami disaster preparedness and mitigation efforts. However, existing approaches based on nonlinear long wave theory, which is commonly used to analyze tsunami propagation and inundation in shallow waters, are computationally expensive, thereby limiting their practical application to probabilistic assessment, which requires many analytical values. In this study, we propose an innovative method to reduce the analytical burden of probabilistic tsunami inundation assessment by building a surrogate model using deep neural networks (DNNs). As input values, we use the slip distribution of an earthquake fault, the initial water level distribution, and the water level distribution over time using linear long wave theory. The results show the possibility of predicting tsunami inundation depths and distributions of inundation with some accuracy directly from the slip distributions of earthquake faults rather than from information on initial water levels and subsequent tsunami water levels. These results indicate that a well-trained machine learning model can predict tsunami inundation depths and distributions without any physical model, which could constitute a breakthrough prediction method. If the probabilistic evaluation of the inundation depth and distribution or the evaluation of uncertainty can be easily performed, local tsunami risk assessment and various disaster countermeasures based on such an assessment can be promoted.

Keywords: tsunami, inundation, probabilistic tsunami hazard analysis, neural network, surrogate model, uncertainty analysis, random fault slip