A novel deep learning-based modelling strategy from image of particles to mechanical properties for granular materials with CNN and BiLSTM

Abstract

It will be practically useful to know the mechanical properties of granular materials by only taking a photo of particles. This study attempts to deal with this challenge by developing a novel deep learning-based modelling strategy. In this strategy, the convolutional neural network (CNN) as image identification algorithm is first used to extract the particle information (particle size distribution PSD and morphology) based on the image of a granular sample, and the bidirectional long short-term memory (BiLSTM) neural network is employed to train the model of reproducing mechanical behaviours and induced fabric evolutions of the sample with corresponding particle information. The datasets of images of samples are generated using discrete element method, and the datasets of mechanical properties together with fabric evolutions are obtained through numerical tests on corresponding samples. As a preliminary attempt, two-dimensional biaxial samples and tests with initially isotropic fabric are considered for the sake of simplicity. The feasibility and reliability of the proposed modelling strategy are evaluated through training and testing. All results indicate that the first part of the model based on CNN is capable of accurately identifying PSD of a granular sample, as well as circularity and roundness of particles, using which as connecting parameters the mechanical behaviours together with induced fabric evolutions of granular materials are subsequently well captured by the second part of the model based on BiLSTM. This study provides a basis and a possible way to obtain immediately particle and packing information, mechanical properties and fabric evolutions by leveraging images of granular materials.