[M-IS01] Particulate Gravity Current

[MIS01-P08] Experimental Verification of a 2D Inverse Model Using a Deep Neural Network

*Seiya Fujishima¹, Hajime Naruse¹ (1.Department of Geology and Mineralogy, Graduate School of Science, Kyoto University)

Keywords:Turbidity Current, Inversion, Neural network

Turbidity currents are sediment gravity flows that support suspended sediment by turbulence and are driven by the density difference with the ambient fluid. Turbidity currents are main process that deliver sediment from shallow to deep sea floor and form large landforms. Recently, it became clear that submarine fans, one of the largest topographic features in deep sea floor, can be classified by their plan-form geometry and degree of developments of distributary channels, which are governed by the hydraulic conditions of forming turbidity currents represented by the densimetric Froude and Rouse numbers. However, this idea was proposed based on numerical experiments, and whether it can be applied to actual submarine fans was not fully examined. Therefore, it is necessary to obtain the hydraulic conditions of turbidity currents to verify this idea. In recent years, the direct observations have exhibited the hydraulic conditions of turbidity currents in the confined topography, such as submarine canyons. However, the hydraulic conditions of ancient turbidity currents cannot be measured directly. Thus, the inverse analysis method was developed to estimate hydraulic conditions from turbidites. In particular, an inverse model using a deep neural network (DNN) was proved effective for estimating hydraulic conditions with high accuracy, which was verified by artificial and flume experimental datasets. However, the previous method employed the 1D layer-averaged model as the forward model, so that it could not reconstruct the behavior of the currents flowing over unconfined topography. Thus, this study aims to develop the horizontal 2D inverse method using the DNN models. The performance of the method was then verified using artificial and flume experimental datasets. First, this study performed numerical simulations of the 2D forward model of turbidity currents to generate training datasets that are composed of combinations of the calculation conditions and the resultant depositional features of turbidites. Here, the flow conditions include the flow velocity, concentration, and height at the flow inlet, and the flow duration. The spatial distributions of the sediment volume-per-unit area of multiple grain-size classes represented the depositional features of a turbidite in this study. The DNNs were trained to learn the relationship between these flow conditions and depositional features, developing the inverse model to predict the flow conditions from the features of turbidites. The trained inverse models were applied to artificial and flume experimental datasets to verify the estimation accuracy. As a result, the inversion results for artificial test datasets marked high accuracy in predicting the flow conditions, although there were some outliers. When the inverse models were applied to the flume experiment data, the hydraulic conditions were estimated reasonably. The relative errors in the flow velocity, concentration, and flow duration were within the range reported in previous studies of 1D inverse models, but the flow height estimation was slightly more erroneous. To further explore the validity of the estimated conditions, the forward model calculation was performed using the estimated experimental conditions. The reproduced bed thickness approximated the measured thickness distribution well, whereas the thin-bedded regions were difficult to reconstruct precisely. In summary, the performance of the 2D inverse models was nearly equivalent to previous 1D models despite their capability of considering complex unconfined topography. The approach used in this study is practical to apply to turbidites in actual submarine fans with the reasonable computational costs and be able to estimate hydraulic conditions with high accuracy. The inverse model developed in this study is expected to contribute to the clarification of the relationship between the morphotype of submarine fans and the hydraulic conditions of turbidity currents.