Application of supervised machine learning as a method for identifying DEM contact law parameters
Calibration of Discrete Element Method (DEM) models is an iterative process of adjusting input parameters such that the macroscopic results of simulations and experiments are similar. Therefore, selecting appropriate input parameters of a model effectively is crucial for the efficient use of the met...
Guardado en:
| Autor principal: | |
|---|---|
| Formato: | article |
| Lenguaje: | EN |
| Publicado: |
AIMS Press
2021
|
| Materias: | |
| Acceso en línea: | https://doaj.org/article/b3810644237b4d258a753edb0674cfb5 |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| Sumario: | Calibration of Discrete Element Method (DEM) models is an iterative process of adjusting input parameters such that the macroscopic results of simulations and experiments are similar. Therefore, selecting appropriate input parameters of a model effectively is crucial for the efficient use of the method. Despite the growing popularity of DEM, there is still an ongoing need for an efficient method for identifying contact law parameters. Commonly used trial and error procedures are very time-consuming and unpractical, especially in the case of models with many parameters to calibrate. It seems that machine learning may offer a new approach to that problem. This research aims to apply supervised machine learning to figure out the dependencies between specific microscopic and macroscopic parameters. More than 6000 DEM simulations of uniaxial compression tests gathered the data for two algorithms - Multiple Linear Regression and Random Forest. Promising results with an accuracy of over 99% give good hope for finding a universal relation between input and output parameters (for a specific DEM implementation) and reducing the number of simulations required for the calibration procedure. Another pertinent question concerns the size of the DEM models used during calibration based on the uniaxial compression test. It has been proven that calibration of certain parameters can be done on smaller samples, where the critical threshold is around 30% of the radius of the original model. |
|---|