11:45 AM - 12:00 PM
[MIS20-09] Estimation of heterogenous source fault model from tsunami deposit by inverse model using DNN
★Invited Papers
Keywords:Tsunami deposit, Inverse model, Fault parameter, 2011 Tohoku-oki tsunami
Accurate estimation of fault parameters of tsunami sources is essential for effective tsunami risk assessment. Complex tsunami behavior can be predicted from the fault parameters of the wave source using the forward model calculation (e.g., Satake et al., 2013); thus, assessing the fault parameters of repeated large tsunamis contributes to disaster prevention planning. However, data on tsunami hydraulics (e.g., wave height and inundation areas), which is used in the current estimation method (e.g., Imamura et al., 2012; Satake et al., 2013), are not available for past tsunamis. Therefore, tsunami deposits are essential sources of information that facilitate the reconstruction of the scale of past tsunamis.
Previously, the deep neural network (DNN) inversion method has been employed to estimate the flow conditions using the 1D forward model (Mitra et al., 2020, 2021; Naruse and Nakao, 2021). The model was applied to actual tsunami deposits of the 2011 Tohoku-oki and the 2004 Indian Ocean tsunamis to predict reasonable values for the hydraulic conditions of the tsunami around the shoreline. In this method, the forward model calculation was repeated with various initial parameters to produce datasets of thickness and grain size distributions of onshore tsunami deposits. Then, the DNN inverse model was trained to learn the relationship between the initial model parameters and the depositional characteristics. In this study, differing from previous studies, the proposed inverse model attempts to estimate the source fault characteristics directly from tsunami deposits by using a 2D forward model, i.e., Delft3D-FLOW (Deltares, 2021), and a DNN.
Numerous fault models have been proposed for the 2011 Tohoku-oki earthquake and implied the complexity of the source fault of this event. Considering the source fault with heterogenous slip is necessary to estimate the tsunami behavior in detail. This study employed the tsunami recipe as a source fault model, which characterizes trench-axis source faults with simplified parameters (HERP, 2017). The fault model is divided into three zones, which are the super large slip (SLS), large slip (LS), and background (BG). The source fault model is defined by four primary parameters: the length, width, and positions of the BG and SLS zones. In total, 728 sets of the source fault primary parameters and tsunami deposit data were generated by the forward model. The moment magnitudes of the fault models were distributed from Mw 7.9 to 9.2
The training with the artificial dataset successfully generated an inverse model of tsunami deposits without overlearning. The test results with 35 sets of artificial data proved the performance of the inverse model. The values predicted by the inverse model agree well with the values given to the original forward model. Among the fault parameters, however, the position of the BG zone does not indicate a precise fitting between the original and estimated values compared with the test results of other parameters. The predicted values of the moment magnitude of the fault model also matched with the original values.
The DNN inverse model was applied to the tsunami deposit formed by the 2011 Tohoku-oki tsunami in three regions: the Sendai Plain in the Miyagi Prefecture, the Odaka Area in the Fukushima Prefecture, and the Rikuzentakata in the Iwate Prefecture. The inverse model predicted the length and width of the source fault of this event as 464.5 km and 162.0 km, respectively. The positions of BG and SLS were estimated to be 196.9 km and 380.9 km from the edge of the Japan Trench, respectively. The moment magnitude was calculated to be 9.0. To examine the validity of the inversion result, the forward model calculation was performed using estimated source fault parameters. The resultant thickness and grain size distribution of the tsunami deposits in each region exhibited the landward fining trend and correlated well with the measured values.
Previously, the deep neural network (DNN) inversion method has been employed to estimate the flow conditions using the 1D forward model (Mitra et al., 2020, 2021; Naruse and Nakao, 2021). The model was applied to actual tsunami deposits of the 2011 Tohoku-oki and the 2004 Indian Ocean tsunamis to predict reasonable values for the hydraulic conditions of the tsunami around the shoreline. In this method, the forward model calculation was repeated with various initial parameters to produce datasets of thickness and grain size distributions of onshore tsunami deposits. Then, the DNN inverse model was trained to learn the relationship between the initial model parameters and the depositional characteristics. In this study, differing from previous studies, the proposed inverse model attempts to estimate the source fault characteristics directly from tsunami deposits by using a 2D forward model, i.e., Delft3D-FLOW (Deltares, 2021), and a DNN.
Numerous fault models have been proposed for the 2011 Tohoku-oki earthquake and implied the complexity of the source fault of this event. Considering the source fault with heterogenous slip is necessary to estimate the tsunami behavior in detail. This study employed the tsunami recipe as a source fault model, which characterizes trench-axis source faults with simplified parameters (HERP, 2017). The fault model is divided into three zones, which are the super large slip (SLS), large slip (LS), and background (BG). The source fault model is defined by four primary parameters: the length, width, and positions of the BG and SLS zones. In total, 728 sets of the source fault primary parameters and tsunami deposit data were generated by the forward model. The moment magnitudes of the fault models were distributed from Mw 7.9 to 9.2
The training with the artificial dataset successfully generated an inverse model of tsunami deposits without overlearning. The test results with 35 sets of artificial data proved the performance of the inverse model. The values predicted by the inverse model agree well with the values given to the original forward model. Among the fault parameters, however, the position of the BG zone does not indicate a precise fitting between the original and estimated values compared with the test results of other parameters. The predicted values of the moment magnitude of the fault model also matched with the original values.
The DNN inverse model was applied to the tsunami deposit formed by the 2011 Tohoku-oki tsunami in three regions: the Sendai Plain in the Miyagi Prefecture, the Odaka Area in the Fukushima Prefecture, and the Rikuzentakata in the Iwate Prefecture. The inverse model predicted the length and width of the source fault of this event as 464.5 km and 162.0 km, respectively. The positions of BG and SLS were estimated to be 196.9 km and 380.9 km from the edge of the Japan Trench, respectively. The moment magnitude was calculated to be 9.0. To examine the validity of the inversion result, the forward model calculation was performed using estimated source fault parameters. The resultant thickness and grain size distribution of the tsunami deposits in each region exhibited the landward fining trend and correlated well with the measured values.