Advancing nearshore and onshore tsunami hazard approximation with machine learning surrogates

Abstract. Probabilistic tsunami hazard assessment and probabilistic tsunami risk assessment (PTHA and PTRA) are vital methodologies for computing tsunami risk and prompt measures to mitigate impacts. However, their application across extensive coastlines, spanning hundreds to thousands of kilometres, is limited by the computational costs of numerically intensive simulations. These simulations often require advanced computational resources, like high-performance computing (HPC), and may yet necessitate reductions in resolution, fewer modelled scenarios, or use of simpler approximation schemes. To address these challenges, it is crucial to develop concepts and algorithms for reducing the number of events simulated and more efficiently approximate the needed simulation results. The case study presented herein, for a coastal region of Tohoku, Japan, utilises a limited number of tsunami simulations from submarine earthquakes along the subduction interface to build a wave propagation and inundation database. These simulation results are fit using a machine learning (ML)-based variational encoder–decoder model. The ML model serves as a surrogate, predicting the tsunami waveform on the coast and the maximum inundation depths onshore at the different test sites. The performance of the surrogate models was assessed using a 5-fold cross-validation assessment across the simulation events. Further, to understand their real-world performance and generalisability, we benchmarked the ML surrogates against five distinct tsunami source models from the literature for historic events. Our results found the ML surrogate to be capable of approximating tsunami hazards on the coast and overland, using limited inputs at deep offshore locations and showcasing their potential in efficient PTHA and PTRA.