Integrating machine learning with DEM for effective calibration of bonded-particle structures

Bonded-particle structures are widely employed to model diverse materials, ranging from brittle ceramics to ductile metals. The mechanical behavior of these structures is fundamentally determined by particle arrangements and bond models representing inter-particle interactions. By maintaining a similar particle arrangement, the structure’s mechanical properties can be tuned through adjustments to bond model parameters. This study introduces a machine-learning (ML) approach for calibrating bond model parameters in bonded-particle structures to match real material properties. We conducted multiple tensile test simulations with varying bond parameters to generate a comprehensive set of stress–strain data, which was subsequently used to train ML models. Various models including neural networks, random forests, K-nearest neighbors, and support vector machines were used, and their performance was evaluated and compared. The trained ML models, in conjunction with the discrete element method (DEM) model, were employed to map the properties of stainless steel AISI 304 onto a bonded-particle solid for a specific set of particle-size distribution, particle arrangement, and loading conditions. To validate the ML approach, we compared the ML-calibrated bond parameters with those obtained through traditional DEM calibration. Our results demonstrate that ML models can effectively capture the complex, high-dimensional relationship between bond parameters and the mechanical response of bonded-particle structures in tensile regime. Given sufficient training data, the calibrated bond model accurately predicts the structure’s mechanical behavior under tensile loading conditions. Although the present study focuses on bond-parameter calibration for a fixed particle arrangement, the proposed framework is readily extendable to investigate the influence of different particle arrangements and other loading conditions through appropriate training datasets. This research establishes a framework for calibrating bond model parameters in bonded-particle structures, offering potential applications in modeling a wide spectrum of materials with complex mechanical behavior. The proposed ML-based approach presents a more efficient and accurate alternative to traditional calibration methods, potentially accelerating materials design and analysis in DEM framework.