Parameter calibration of the angle of repose of particle materials based on convolutional neural network

Accurate determination of microscopic parameters is crucial for employing the discrete element method in addressing practical engineering challenges. The angle of repose calibration method for bulk materials is employed but frequently relies on subjective human measurements, potentially resulting in errors. This paper introduces a parameter calibration method that utilises a convolutional neural network to enhance standardisation, universality, and accuracy in predicting particle material behaviour. Firstly, the angle of repose simulations are conducted to establish training and test datasets. Next, sensitivity analysis is performed to determine the evaluation index. Subsequently, the performance differences in prediction accuracy among various input data types and network models, including one-dimensional convolutional, two-dimensional convolutional, and fully connected networks were compared. Finally, the influence of particle size and material type on the trained network model was investigated. The experimental results demonstrate that convolutional neural networks outperform traditional parameter calibration methods, in terms of feature extraction capabilities. According to the evaluation indicators in this paper, the conventional method achieves the highest prediction accuracy of 63.33%, whereas the deep learning method achieves a prediction accuracy of 86.67%. Additionally, the accuracy of one-dimensional convolutional network predictions is relatively high when compared to two-dimensional convolutional and fully connected networks. Furthermore, contour feature data exhibits superiority over slope data. Specifically, when the network input data consists of contour data, the prediction accuracy is further enhanced by 6.67% due to its inclusion of more effective features. This study provides new insights into the angle of repose parameter calibration.