A discrete element modeling of rock and soil material based on the machine learning

Abstract

The discrete element modeling of rock and soil mass usually relies on tedious parameter adjustment and mechanical property testing operations to obtain appropriate mechanical parameters of the element. In this paper, based on the linear elastic contact theory, a discrete element model of rock samples was established based on discrete element software MatDEM with high-performance numerical analysis, and the actual mechanical properties of the model samples were predicted using a machine learning method, XGBoost algorithm. A certain amount of sample data sets was generated, and through numerical experiments, training, validation and test operation of the rock mechanical properties (compressive strength, Young's modulus, Poisson's ratio and tensile strength), the numerical model quantitative relationship between the actual and input mechanical properties was explored, showing the following results: (1) Based on the characteristics of the model 12000 rock samples and correlation analysis, the tag data shows that the input mechanical properties are positively correlated with the actual mechanical properties;(2) With the increase of the number of samples in the sample training data set, the training average error becomes smaller and smaller, and the error distribution becomes more and more stable.(3) The training results converge to the set threshold, and the prediction errors of mechanical parameters are all less than 4%, and the error distribution is stable, which proves that the machine learning method can quickly and accurately establish the rock model with specific mechanical properties.