Deep learning long short-term memory combined with discrete element method for porosity prediction in gravel-bed rivers

• Published in: International journal of sediment research 2023-02, Vol.38 (1), p.128-140
• Main Authors: Anh, Duong Tran, Tanim, Ahad Hasan, Kushwaha, Daniel Prakash, Pham, Quoc Bao, Bui, Van Hieu
• Format: Article
• Language: English
• Subjects: Bed porosity; Discrete element method; Gravel-bed river; LSTM; Numerical modeling
• ISSN: 1001-6279
• DOI: 10.1016/j.ijsrc.2022.08.001

The porosity of gravel riverbed material often is an essential parameter to estimate the sediment transport rate, groundwater-river flow interaction, river ecosystem, and fluvial geomorphology. Current methods of porosity estimation are time-consuming in simulation. To evaluate the relation between porosity and grain size distribution (GSD), this study proposed a hybrid model of deep learning Long Short-Term Memory (LSTM) combined with the Discrete Element Method (DEM). The DEM is applied to model the packing pattern of gravel-bed structure and fine sediment infiltration processes in three-dimensional (3D) space. The combined approaches for porosity calculation enable the porosity to be determined through real time images, fast labeling to be applied, and validation to be done. DEM outputs based on the porosity dataset were utilized to develop the deep learning LSTM model for predicting bed porosity based on the GSD. The simulation results validated with the experimental data then segregated into 800 cross sections along the vertical direction of gravel pack. Two DEM packing cases, i.e., clogging and penetration are tested to predict the porosity. The LSTM model performance measures for porosity estimation along the z-direction are the coefficient of determination (R2), root mean squared error (RMSE), and mean absolute error (MAE) with values of 0.99, 0.01, and 0.01 respectively, which is better than the values obtained for the Clogging case which are 0.71, 0.14, and 0.03, respectively. The use of the LSTM in combination with the DEM model yields satisfactory results in a less complex gravel pack DEM setup, suggesting that it could be a viable alternative to minimize the simulation time and provide a robust tool for gravel riverbed porosity prediction. The simulated results showed that the hybrid model of the LSTM combined with the DEM is reliable and accurate in porosity prediction in gravel-bed river test samples.