Modelling of rheological characteristics of glacier debris flow using artificial neural networks

In this study, 52 debris flow deposits along the Sichuan Tibet railway were collected through field investigation. Subsequently, rheological analysis was carried out for each debris flow in the laboratory, and the stress and strain characteristic parameters of the debris flow were obtained. The main objective of this study was to establish a neural network to predict the shear stress of debris flow. The backpropagation neural network (BTNN), recurrent neural network (RNN) and general regression neural network (GRNN) were used to train the input data (debris flow gradation, bulk density of debris flow, shear rate) and output data (debris flow shear stress), which were obtained from the grading analysis and rheological analysis of the debris fluid in the laboratory. The predicted results reveal that the root mean square error of the three neural networks is 0.007 (GRNN), 0.018 (RNN) and 0.019 (BTNN), which demonstrates the good prediction stability of the neural networks. The RNN model has an excellent R^2 value (0.95), which is better than that of the GRNN (0.93) and BTNN models (0.91). Overall, the RNN neural network has the highest prediction accuracy, and the order of performance for the developed architectures is RNN > GRNN > BPNN, based on their comparison with the test data. The comparison of the optimized ANN model with conventional multiple linear regression (MLR) reveals that the ANN model achieved substantially better prediction performance compared with the MLR models. The RNN neural network established in this study provides a new method for more quickly and accurately predicting the shear stress value of debris flow. Finally, the prediction results obtained by the neural network were compared with those obtained by a traditional rheological model, and it was found that the prediction results of the neural network are better than those obtained by the traditional rheological model.