Optimization of DEM parameters using multi-objective reinforcement learning

• Published in: Powder technology 2021-02, Vol.379, p.602-616
• Main Authors: Westbrink, Fabian, Elbel, Alexander, Schwung, Andreas, Ding, Steven X.
• Format: Article
• Language: English
• Subjects: Bulk handling; Calibration; Discrete element method; Discrete element method (DEM); Learning; Material properties; Multi-objective; Multiple objective analysis; Optimization; Parameter calibration; Parameters; Reinforcement; Reinforcement learning
• ISSN: 0032-5910; 1873-328X
• DOI: 10.1016/j.powtec.2020.10.067

Simulations with the Discrete Element Method (DEM) have become prominent for analyzing bulk behavior in various industries. For each application the material has to be analyzed while the material parameters have to be determined to ensure a valid and reliable result. However, material properties available in the literature are hardly usable and unsuitable for a macroscopic analysis of the bulk behavior. Thus, the material has to be tested and evaluated to calibrate it with suitable DEM material parameters. In this work, a novel approach for DEM calibration with a parameter optimization based on multi-objective reinforcement learning is proposed. This approach uses the results of two different environments and trains an agent to find a suitable material parameter-set with a low number of required iterations and a small number of hyper-parameters. To ensure the applicability of the developed approach, three materials with different characteristics are calibrated and validated. [Display omitted] •Novel DEM optimization procedure using multi-objective reinforcement learning•Remarkable optimization performance with a low number of required DEM simulations•Pre-training procedure results in a highly generalizing agent which can be applied to arbitrary materials.•Calibration and validation of three different materials (cohesionless up to slightly cohesion behavior)