A Machine Learning Model of Liquid Bridge Force and Its Application in Discrete Element Method

Hanwen Cui,Honghua Zhao,Shunying Ji,Xingli Zhang,Waleed Awadalseed,Hongxiang Tang
DOI: https://doi.org/10.1016/j.conbuildmat.2023.134174
IF: 7.4
2024-01-01
Construction and Building Materials
Abstract:Developing an accurate and efficient contact model for discrete element method (DEM) simulation that can describe the microscopic dynamics of low moisture content wet granular systems continues to be a major challenge. In this study, a improved energetic approach based on the traditional linear unit discretization method is introduced to solve liquid bridge problems. This improved approach significantly ameliorates the efficiency of constructing the databases while ensuring high data accuracy. Using this method, we generated databases consisting of 163,440,000 and 2043,000 data points for the sphere-sphere and sphere-wall systems, respectively. Machine learning capillary force models (ML-Capillary) are trained using the Catboost algorithm. These models demonstrate excellent performance in predicting capillary force between particles as well as between particles and walls, considering different particle sizes, moisture levels, and contact angles, while neglecting the influence of gravity. The corresponding relative errors are below 1.5% and 2%, respectively, which are lower than existing simulation models. We incorporated the obtained models into the discrete element method and calculated capillary force between particle pairs and simulated wet granular column collapse tests. The simulation results show good agreement with experimental results in terms of parameters such as inter-particle capillary forces and particle column collapse run-out length. This study will provide an accurate and efficient numerical model for simulating low saturation wet granular systems using the discrete element method.
What problem does this paper attempt to address?