Subsequently, the calibrated numerical model was implemented for lateral intake at 31 positions of the external bend from the section of 10 to 140 degrees of the bend with 5-degree intervals and with 5 diversion angles of 10, 30, 50, 70, and 90 degrees and three intake discharge ratios of 20%, 30%, and 40%, and the percentage of sediment diverted to the lateral intake was determined for each model. The results were used as necessary data for training and validation of various neural network models. Next, by extracting the results, using 70% of the data, a meta-model was trained using the multi-layer artificial neural network (MLP), radial basis function network (RBF), generalized regression neural network (GRNN), and support vector regression (SVR) models, and its accuracy was evaluated using a part of the data that was not used in the training phase. The accuracy of the results of the aforementioned model in terms of the error index of the mean absolute magnitude of the relative error in the validation period (3.6%) indicates the better ability of the SVR neural network model in predicting the amount of sediment input to the lateral intakes.
Materials and methods
- Laboratory and numerical model
In the present study, first, using a discrete phase numerical model, the sediment transport phenomenon in a 180-degree curved channel with a lateral intake was simulated, and the sediment transport mechanism and the amount of sediment entering the intake under different hydraulic conditions were investigated (Tavakoli et al., 2019). To verify the numerical model, the results of the laboratory study by Montaseri (2008) were used. This model includes a U-shaped channel with an average radius of 2.6 m and a width of 0.6 m.
- Artificial Neural Network Models
Artificial neural networks, inspired by the human body's neural network, are composed of components called neurons and include the following:
- Multilayer Neural Network (MLP) model
- Radial Basis Function (RBF) networks
- Generalized Regression Neural Network (GRNN)
- Support Vector Regression (SVR)

Results and discussion
The analysis of different artificial neural network models in predicting the amount of sediment entering the lateral catchment in a U-shaped arched channel reveals that the support vector regression (SVR) model achieved the highest correlation coefficient (R² = 0.982) between observed and predicted data. The radial basis function (RBF) model followed closely with an R² value of 0.98, while the multilayer perceptron (MLP) and general regression neural network (GRNN) models exhibited R² values of 0.976 and 0.949, respectively.

Conclusion
In this study, based on the data obtained from the calibrated numerical model, a metamodel was used in different neural network models to predict the amount of sediment entering the lateral intakes in the U-shaped arched channel, and different architectures of neural network models were examined. The results show that the best multilayer neural network architecture with two hidden layers and 5 neurons in the first hidden layer and 5 neurons in the second hidden layer was obtained. The best RBF neural network architecture was obtained with a Spread value of 0.25. The best GRNN neural network architecture was obtained with a Spread value of 5. The results show that the SVR model has the maximum correlation coefficient (R2) with the observational data. Therefore, it is recommended to use the support vector regression model to predict the sediment entering the lateral intakes.
In this study, a predictive metamodel based on the Fluent numerical model and various artificial neural network models is presented to estimate the amount of sediment diverted to lateral catchments located in river bends based on input data including the catchment location, catchment deviation angle, and percentage of inflow. In the first step, using the discrete phase two-phase model in Fluent, the sediment transport phenomenon was simulated in a 180-degree curved channel with a lateral catchment located at a position of 115 degrees from the outer bend and with a deviation angle of 45 degrees, and the results were calibrated with laboratory data. Subsequently, the calibrated numerical model was implemented at 31 positions of the outer bend from a section of 10 to 140 degrees of the bend with 5-degree intervals and with 5 intake angles of 10, 30, 50, 70, and 90 degrees and three intake discharges of 20%, 30%, and 40%, and the percentage of sediment diverted to the lateral catchment was determined for each model. The results were used as necessary data for training and validation of various neural network models. Next, by extracting the results, using 70% of the data, a meta-model was trained using the models of multilayer artificial neural network (MLP), radial branch neural network (RBF), generalized regression neural network (GRNN), and support vector regression (SVR), and its accuracy was evaluated using a part of the data that was not used in the training phase. The accuracy of the results of the aforementioned model in terms of the error index of the mean absolute magnitude of the relative error in the validation period (3.6%) indicates the better ability of the SVR neural network model in predicting the sediment input to the lateral catchment.
In this study, a predictive metamodel based on the Fluent numerical model and various artificial neural network models is presented to estimate the amount of sediment diverted to lateral catchments located in river bends based on input data including the catchment location, catchment deviation angle, and percentage of inflow. In the first step, using the discrete phase two-phase model in Fluent, the sediment transport phenomenon was simulated in a 180-degree curved channel with a lateral catchment located at a position of 115 degrees from the outer bend and with a deviation angle of 45 degrees, and the results were calibrated with laboratory data. Subsequently, the calibrated numerical model was implemented at 31 positions of the outer bend from a section of 10 to 140 degrees of the bend with 5-degree intervals and with 5 intake angles of 10, 30, 50, 70, and 90 degrees and three intake discharges of 20%, 30%, and 40%, and the percentage of sediment diverted to the lateral catchment was determined for each model. The results were used as necessary data for training and validation of various neural network models. Next, by extracting the results, using 70% of the data, a meta-model was trained using the models of multilayer artificial neural network (MLP), radial branch neural network (RBF), generalized regression neural network (GRNN), and support vector regression (SVR), and its accuracy was evaluated using a part of the data that was not used in the training phase. The accuracy of the results of the aforementioned model in terms of the error index of the mean absolute magnitude of the relative error in the validation period (3.6%) indicates the better ability of the SVR neural network model in predicting the sediment input to the lateral catchment.