AI-based ensemble tsunami inundation forecasting

In addition to accuracy and speed, tsunami inundation forecasts should also account for the uncertainty primarily attributed to the earthquake slip distribution, particularly for near-field events. We propose a method based on artificial intelligence (AI) to timely forecast the tsunami inundation from megathrust earthquakes and efficiently quantify such uncertainties. We trained the AI-based tsunami model using 630 hypothetical scenarios representing a wide range of plausible tsunami occurrences in Pacitan regency, southern coast of Java, Indonesia. In the inference stage, the AI model requires inputs of synthetic tsunami waveforms at specified locations obtained through a Green’s function summation technique. Such an approach enables the rapid generation of tsunami waveforms from the corresponding stochastic slip models. Subsequently, the AI model runs multiple ensemble tsunami inundation forecasts derived from its varying set of inputs. The final product of the AI-based ensemble model comprises multiple potential tsunami inundation results as the basis for probabilistic forecasts, and from which the forecast uncertainty can be explicitly defined. A test on an anticipated event generated by an earthquake off the south coast of Java of Mw 8.8 produces an inundation forecast accuracy of 92% using the median ensemble forecast, whereas the standard method based on a uniform slip yields 85%. However, it is noted that the effects of depth-dependent rigidity and tsunami earthquakes are not yet included.