Machine Learning Approaches for Tsunami Hazard and Risk Assessment

Probabilistic workflows are indispensable for assessing the overland tsunami hazard and risk, due to the infrequency and limited historical observations of tsunamis. However, these workflows are computationally demanding because they require a large number of simulations to capture uncertainty of the phenomena. This study leverages machine learning (ML) emulators to address this challenge by directly predicting hazard and risk metrics, bypassing the need for extensive numerical simulations for the inundation phase.

The ML emulators are trained to predict high-resolution hazard metrics onshore (e.g., maximum inundation depth) and risk metrics (e.g., expected damage or loss) using offshore waveforms and local deformation fields as inputs. A database of tsunamigenic earthquakes in the Mediterranean Sea, reflecting substantial variability in source mechanisms and locations, was used for training and validation. For a test site in Sicily, Italy, the emulator demonstrated robust performance with a training set of ~1,600 events, achieving a 30-fold reduction in computational cost compared to traditional probabilistic tsunami hazard assessment (PTHA) workflows.

In the aftermath of tsunami event, such ML emulators can be used to directly provide rapid estimates on the expected damage and losses at different disaggregation, while evaluating many different scenarios due to the uncertainty in the characterization of the earthquake source in the early stages of after the earthquake event.