[M-IS15] 津波堆積物：東北地方太平洋沖地震後10年の成果と今後の展望

[MIS15-04] Application of DNN inverse model to large scale tsunamis to reconstruct the flow conditions of modern and ancient tsunamis

★Invited Papers

*Rimali Mitra¹、Hajime Naruse¹、Tomoya Abe²、Shigehiro Fujino³、Daisuke Sugawara⁴ (1.Kyoto University, Graduate School of Science、2.Research Institute of Geology and Geoinformation, Geological Survey of Japan, National Institute of Advanced Industrial Science and Technology (AIST), Tsukuba, Japan、3.University of Tsukuba、4.Tohoku University, IRIDeS)

キーワード：Tsunami Deposit, DNN inverse model, Machine learning

The 2011 Tohoku-oki tsunami and 2004 Indian Ocean tsunami caused significant economic losses and fatalities in the coastal areas. These modern tsunamis could be traced back and linked to the ancient tsunamis such as 869 Jogan tsunami as a precursor for 2011 Tohoku-oki tsunami. The Deep Neural network (DNN) 1D inverse model has the potential to predict the values of hydraulic conditions such as maximum inundation distance, flow velocity and maximum flow depth, sediment concentration of different grain-size classes using the thickness and grain-size distribution of the tsunami deposit from the post-tsunami survey around Sendai plain and other locations. The DNN model incorporates the forward model, which was based on non-uniform, unsteady transport of suspended sediments during turbulent mixing. The forward model employs iterative random flow conditions as boundary conditions to produce artificial training data sets which represent depositional characteristics, such as, thickness and grain-size distribution. Then the DNN was trained to find relationships between simulation results and initial conditions using the artificial output datasets of forward model. Finally, the DNN model was applied to natural field data sets from different locations. The quantification of uncertainty of the model predicted results was also reported using the jackknife method. The model successfully reconstructed the flow conditions of 2011 Tohoku-oki tsunami from Sendai plain and Joban coast, 2004 Indian Ocean tsunami from Phra Thong island, Thailand and 869 Jogan tsunami from Sendai plain. The estimated of values maximum inundation distance flow velocity, maximum flow depth were reasonable and close to the measured or reported values from the study area. Despite having topographical complicacy and lesser sampling locations for ancient deposits, the DNN model have the potential to predict the flow characteristics from the deposits satisfactorily. These qualitative and quantitative estimations demonstrated that the DNN inverse model is a potential tool for evaluating tsunami hazard and mitigation strategies.