UC Davis

Tsunamis in the last two decades have resulted in the loss of life of over 200,000 people and have caused billions of dollars in damage. Therefore there is great motivation for the development and improvement of current tsunami warning systems. However, this is quite difficult as tsunamigenic sources such as earthquakes are inherently unpredictable due to their non-linear and chaotic nature. As a result, forecasting methods focus on taking advantage of the short span of time between the earthquake rupture and the wave reaching the shore. A common method includes forward modeling, which first obtains the earthquake source by using inversion techniques. It is then used as the initial condition for the tsunami and is simulated in real time to obtain a forecast of the coast. However, methods such as these still take an appreciable amount of time to run and rely on supercomputers which may not be available to countries or institutions which do not have sufficient resources and infrastructure. Therefore, in response to the shortcomings of current methods, other methods have been explored utilizing neural networks trained on precomputed tsunami databases to make inundation forecasts. From the nature of neural networks, these methods obviate the need for supercomputers and cut down the forecasting computation time significantly. The work presented in this dissertation represents advancements made towards the creation of a neural network-based tsunami warning system which can produce faster inundation forecasts with increased accuracy. This was done by first improving the waveform resolution and accuracy of Tsunami Squares, an efficient cellular automata approach to wave simulation. It was then used to create a database of precomputed tsunamis in the event of a magnitude 9+ rupture of the Cascadia Subduction Zone, located only ~100 km off the coast of Oregon, US. Two methods were tested in the effort to link readily available sensor data directly to inundation forecasts. One approach utilized a convolutional neural network which took wave height data from buoys as input and proved successful as maps of maximum inundation could be predicted for the town of Seaside, Oregon with a median error of ~0.5 m.