Identification of representative earthquakes for probabilistic tsunami hazard analysis (PTHA) using earthquake rupture forecasts and machine learning

As probabilistic tsunami hazard analysis (PTHA) focuses more on assessments for localized, populous regions, techniques are needed to identify a subsample of representative earthquake ruptures to make the computational requirements for producing high-resolution hazard maps tractable. Moreover, the greatest epistemic uncertainty in seismic PTHA is related to source characterization, which is often poorly defined and subjective. We address these two salient issues by applying streamlined earthquake rupture forecasts (ERFs), based on combinatorial optimization methods, to an unsupervised machine learning workflow for identifying representative ruptures. ERFs determine the optimal distribution of a millennia-scale sample of earthquakes by inverting the observed slip rate on major faults. We use two previously developed combinatorial optimization ERFs, integer programming and greedy sequential, to produce the optimal location of ruptures with seismic moments sampled from a regional Gutenberg–Richter magnitude–frequency distribution. These ruptures in turn are used to calculate peak nearshore tsunami amplitude, using computationally efficient tsunami Green's functions. An unsupervised machine learning workflow is then used to identify a small subsample of the earthquakes input to ERFs for onshore PTHA analysis. We eliminate epistemic uncertainty related to source distribution under traditional PTHA analysis; in its place, a quantifiable, less subjective and generally smaller uncertainty related to the input to ERFs is included. The Nankai subduction zone is used as a test case, where previous ERFs have been conducted. Results indicate that the locations of representative earthquakes are sensitive to choice of magnitude–area relation and to whether a minimum cumulative stress objective is imposed on the fault. In general, incorporating ERFs into PTHA provide a physically self-consistent method to incorporate fault slip information in determining representative earthquakes for onshore PTHA, eliminating a major source of epistemic uncertainty.