Convolution neural network analysis of the effect of gradation on the mechanical properties of granular materials

<p indent="0mm">Gradation, a fundamental physical parameter of granular materials, considerably influences their mechanical properties. To rapidly obtain mechanical strength indices for particle assemblies with varying continuous gradation distributions and to analyze the mechanisms through which gradation affects mechanical properties, 100 gradation curves representing 3 typical continuous grading distributions were established based on the grading equation. Numerical samples corresponding to these curves were generated using the discrete element method, followed by biaxial compression testing to construct a dataset for training deep neural networks. Bayesian optimization methods were used to determine the optimal network hyperparameters and enhance the predictive performance of the model. Using compressed numerical sample images and their corresponding mechanical strength as training data, GoogLeNet was used to predict the shear strength of granular samples with varying gradations. By adjusting the input image information, this study examined the effects of particle and pore distribution, particle size, high-stress particles, and force chains on the predictive performance of the model. The results indicated that the convolutional neural network effectively captured the relationship between gradation characteristics and the mechanical properties of granular materials. Notably, information related to high-stress particles and force chains emerged as critical factors influencing the mechanical properties of granular materials and the predictive accuracy of the neural network model.