Stochastic Emulation for Efficient Onshore Probabilistic Tsunami Hazard Assessment
Journal
JOURNAL OF GEOPHYSICAL RESEARCH. MACHINE LEARNING AND COMPUTATION
ISSN
2993-5210
2993-5210
Date Issued
2026
Author(s)
Ramalingam, Naveen
Storrøsten, Erlend Briseid
DOI
10.1029/2025JH001100
Abstract
Machine learning is emerging as a promising strategy for modeling tsunami inundation at reduced computational cost. To enable probabilistic outputs that capture emulator uncertainty, we employ an ensemble of stochastic encoder-decoder emulators with dropout-based stochastic forward pass. Training and testing is conducted on an extensive Mediterranean Sea tsunami simulation data set comprising multiple earthquake source regions with variable magnitude, location, and focal mechanism, including near-field events with local coseismic deformation. For the two target sites Catania and Siracusa in Sicily, we achieve good performance for relatively sparse training sets of only a few hundred events. We demonstrate the potential of our emulators to accurately model inundation from a range of tsunami sources needed to perform probabilistic tsunami hazard assessment (PTHA), reducing the computational cost of PTHA by a factor of 30-100, and estimating the uncertainty introduced by the emulator. Comparing probabilistic hazard maps and curves from our emulation with those from the full simulation ensemble and importance sampling Monte Carlo approach, we demonstrate the efficacy of emulators in capturing the spatial distribution and magnitude of inundation needed in overland tsunami hazard assessment. Plain Language Summary Probabilistic Tsunami Hazard Analysis (PTHA) requires simulating tens of thousands of tsunami scenarios, with high-resolution inundation modeling being the most computationally expensive step. Machine learning (ML) offers a way to predict inundation outcomes using only a limited number of full simulations for training, greatly reducing computational cost. In this study, we develop an ML model to predict detailed tsunami inundation maps from offshore wave data and local ground deformation caused by earthquakes. The model is trained on simulated Mediterranean tsunamis and tested at two sites in Eastern Sicily, using earthquakes that generate offshore tsunami waves and, in some cases, local ground shifts. Despite using only a few hundred training events, in testing it accurately predicted flood depths across a wide range of earthquake scenarios. The approach not only estimates expected inundation but also provides a range of possible outcomes and their likelihoods, improving uncertainty quantification. By incorporating ML into PTHA, we can efficiently evaluate thousands of potential tsunami events, making large-scale hazard and risk assessment feasible. This advancement enables more cost-effective and timely analysis, supporting better disaster preparedness, planning, and resilience for coastal communities exposed to tsunami hazards.
