Keywords:Tsunami Deposit, Inverse Model, 2011 Tohoku-Oki Tsunami, Deep Learning Neural Network
Tsunami deposits provide clues for estimating the magnitude and hydraulic conditions of paleo-tsunamis, which can serve as a tool for evaluating tsunami hazard and mitigation strategies. In this study, a numerical approach towards inverse modelling of tsunami deposit has been taken. The present model uses Deep learning Neural Network (DNN) to understand quantitative hydraulic conditions of tsunamis from the tsunami deposits using non-uniform, unsteady transport of suspended sediments during turbulent mixing. This 1D inverse model incorporates a forward model which employs iterative random flow conditions as boundary conditions (e.g. maximum flow depth, flow velocity, concentration, etc.) to produce artificial training data sets which represent depositional characteristics, such as, thickness and grain-size distribution. The machine learning of DNN includes several hyperparameters which are used for an optimizer (Stochastic Gradient Descent) and activation function (Rectified Linear Unit). Subsequently, DNN was trained to find relationships between simulation results and initial conditions using the artificial output datasets of forward model. The uncertainties in the model were calculated using Jackknife method. Finally, the inverse model was applied to natural field data sets from 2011 Tohoku-Oki Tsunami deposits around the Sendai plain and Odaka in Joban coastal plain, exhibiting that the method can reconstruct hydraulic conditions of paleo-tsunamis. The proposed inverse model in tandem with other forward models such as Delft3D are expected to simulate unsteady 2D hydrodynamics and sediment transport. Thus, in future studies, it is expected that this model could be used to successfully reconstruct the flow conditions of modern and older tsunamis with the uncertainty analysis.