Sequential Bayesian Update to Detect the Most Likely Tsunami Scenario Using Observational Wave Sequences.

This study presents a method for the detection of the most likely tsunami scenario among a set of possible scenarios using an observational wave sequence based on a sequential Bayesian update scheme. The proposed method consists of two phases: an offline preliminary learning phase and an online real‐time detection update phase. The innovation of this study is that proper orthogonal decomposition (POD) and Bayesian update are used together with an established tsunami simulation technique. In the offline reinforcement learning process, a series of tsunami simulations are carried out based on geophysically feasible scenarios, and the spatial modes of wave data calculated at predefined synthetic gauge locations are extracted through the application of POD. When a real tsunami event occurs and observational ocean data are obtained, the online process can then be performed as follows: using the stored spatial modes along with their component coefficients, pseudocoefficients are repeatedly estimated from the obtained wave data and used to sequentially update the most likely tsunami scenario according to the posterior probability through Bayesian update. A verification analysis is carried out to illustrate the procedure of the proposed method, and a validation analysis is conducted to demonstrate both the capabilities and applicability of the process with reasonable accuracy. A comprehensive discussion details the characteristic features of the proposed method in terms of the real‐time prediction of tsunami hazards and risks. Plain Language Summary: We present a real‐time tsunami scenario detection framework using sequential probability updates and a kind of unsupervised learning. We first carry out a series of tsunami simulations based on geophysically feasible earthquake scenarios. Then, the characteristics of simulated wave history data from predefined ocean gauges are learned by means of proper orthogonal decomposition (POD), that is, the commonalities ("spatial modes") among all tsunami scenarios are extracted. The scenario‐specific components ("component coefficients") obtained together with such spatial modes enable us to handle all learned scenarios. When a real tsunami event occurs, the extracted spatial modes are used to infer the specific components of the current event, called "pseudocoefficients", from the real‐time wave observations. By inputting these "pseudocoefficients" and the prelearned "component coefficient" into a Bayesian update scheme, the likelihood that each of the learned scenarios corresponds to the current event is sequentially evaluated. To demonstrate the specific procedure and its capabilities, validation analyses are conducted targeting the Nankai subduction zone. Our framework successfully detects the most likely scenarios, which have wave histories, maximum wave heights, and fault rupture patterns similar to those of the test scenarios, from the training dataset. Key Points: A sequential Bayesian update scheme for the detection of the most likely tsunami scenario using proper orthogonal decomposition is presentedTwo case studies targeting the Nankai Trough are carried out to demonstrate the procedure and capabilities of the proposed methodThe most likely tsunami scenarios are detected with 7 min of observational wave data from 71 synthetic gauge points [ABSTRACT FROM AUTHOR]