4:30 PM - 4:45 PM
[MIS15-04] Development of 2-D inverse model for turbidity currents based on deep learning
Keywords:Turbidity current, Turbidite, Machine learning, Deep learning, Inverse analysis
A turbidity current is a sediment gravity flow that supports suspended sediment by turbulence and is driven by the density difference with the ambient fluid. In recent years, in-situ observations of turbidity currents in submarine canyons have revealed hydraulic conditions such as flow velocity or depth in confined topographic situations. However, laterally expanding turbidity currents have not been directly observed, and the hydraulic conditions of such currents remain unknown. Turbidity currents generated by earthquakes and tsunamis are known to flow over a wide area, so elucidation of the hydraulic conditions of laterally spreading turbidity currents is critical to estimating the magnitude of past geologic events. Therefore, we developed a horizontal 2-D inverse model to estimate the hydraulic conditions of turbidity currents from turbidites. The model performance was verified[hn1] using the results of the flume experiment.
The procedures for developing the inverse model and verifying model performance are following. First, the flow conditions (the flow velocity, suspended sediment concentration, flow depth at the inlet, and flow duration) were randomly generated, and those conditions were given to the forward model to produce the thickness and grain size distributions of turbidites. Combinations of those flow conditions and turbidite features were used as a training dataset of the inverse model. After that, a fully connected neural network learned the relationship between flow conditions and turbidite characteristics. As a result of this training process, the inverse model that returns flow conditions based on turbidite thickness and grain size distribution was developed. Artificial test datasets tested the inverse model. Also, the inverse model was applied to the flume experimental turbidites to test the model's performance. As a result, the flow conditions (velocity, suspended sediment concentration, flow depth at the upstream end boundary, and flow duration) were accurately reconstructed from artificial test datasets. Moreover, the flow behavior predicted by the estimated conditions matched with the true behaviors. For the experimental-scale turbidity currents, the flow depth, suspended sediment concentration for grain-size classes exhibiting relatively high values, and flow duration showed small relative errors between predicted and measured values. However, the flow velocity and suspended sediment concentration of other grain size classes were overestimated. The erroneous predictions of suspended sediment concentration were considered to be a measurement error because they were below the measurement limit of the electronic balance. The error of the flow velocity was probably caused because the basal friction coefficient was too small to produce the bypass zones in the upstream region of the experimental flume. The high flow velocity was required to reproduce such channel-like topography at the low value of the friction coefficient. In the future, the parameters of the forward model will be adequately determined to improve the inversion accuracy.
The procedures for developing the inverse model and verifying model performance are following. First, the flow conditions (the flow velocity, suspended sediment concentration, flow depth at the inlet, and flow duration) were randomly generated, and those conditions were given to the forward model to produce the thickness and grain size distributions of turbidites. Combinations of those flow conditions and turbidite features were used as a training dataset of the inverse model. After that, a fully connected neural network learned the relationship between flow conditions and turbidite characteristics. As a result of this training process, the inverse model that returns flow conditions based on turbidite thickness and grain size distribution was developed. Artificial test datasets tested the inverse model. Also, the inverse model was applied to the flume experimental turbidites to test the model's performance. As a result, the flow conditions (velocity, suspended sediment concentration, flow depth at the upstream end boundary, and flow duration) were accurately reconstructed from artificial test datasets. Moreover, the flow behavior predicted by the estimated conditions matched with the true behaviors. For the experimental-scale turbidity currents, the flow depth, suspended sediment concentration for grain-size classes exhibiting relatively high values, and flow duration showed small relative errors between predicted and measured values. However, the flow velocity and suspended sediment concentration of other grain size classes were overestimated. The erroneous predictions of suspended sediment concentration were considered to be a measurement error because they were below the measurement limit of the electronic balance. The error of the flow velocity was probably caused because the basal friction coefficient was too small to produce the bypass zones in the upstream region of the experimental flume. The high flow velocity was required to reproduce such channel-like topography at the low value of the friction coefficient. In the future, the parameters of the forward model will be adequately determined to improve the inversion accuracy.