Convolutional neural networks for tsunami intensity measure prediction from slip and coseismic deformation data

Subduction zones can produce catastrophic tsunamis. In recent years, Chile has experienced destructive tsunamigenic earthquakes, including the February 27, 2010, with a magnitude of 8.8, another of 8.2 on April 1, 2014, and 8.3 on September 16, 2015. Accurately predicting Tsunami Intensity Measures (TIMs) is crucial for assessing tsunami hazard and mitigating risk. Despite the frequency of these events, observational data remains scarce, and numerical tsunami inundation simulations are expensive. To address these challenges, we propose a novel method using two-dimensional convolutional neural networks combined with fully connected nodes to predict key TIMs: run-up, inundation area, and wave height. Our approach is applied in four cities in Central Chile: Viña del Mar, Valparaíso, Cartagena, and San Antonio. We develop and compare two models that differ in their input: one uses seismic slip images, and the other uses coseismic deformation images. To assess their performance, we compare them with literature architectures, such as Visual Geometry Group 16 (VGG16) and Residual Network 50 (ResNet50). Our model using coseismic deformation demonstrates superior accuracy, achieving coefficient of determination (R²) values between 0.79 and 0.98 across all TIMs, compared to 0.14 to 0.91 for VGG16 and 0.19 to 0.81 for ResNet50. This improvement is attributed to the stronger correlation of coseismic deformation in the physical processes driving tsunami generation. To provide a probabilistic framework for assessing the reliability of our predictions, we quantify uncertainties using Bayesian-based models. This method offers an efficient alternative to costly simulations, enhancing tsunami hazard assessments and tsunami-resilient structural design.