Maschinelles Lernen für die numerische Homogenisierung von Beton

authored by: Fadi Aldakheel, Michael Haist, Ludger Lohaus, Peter Wriggers
Abstract: Material modeling of concrete using modern numerical methods significantly accelerates the design process of structures. However, for multiscale modeling of such a heterogeneous material, the established homogenization methods are still very computationally intensive, especially for high accuracy requirements. In this paper, we propose a machine learning approach that provides a computationally efficient solution method while delivering a high degree of accuracy. The dataset used for the training and testing process consists of artificial and real microstructural images (input), while the result data (output) are the homogenized stresses of a given representative volume element (RVE). The performance of the model is demonstrated by examples and compared with classical homogenization methods. The developed ML model achieves higher accuracy in determining the homogenized stresses and significantly reduces the computation time.
Organisation(s): Institute of Mechanics and Computational Mechanics
Institute of Building Materials Science
Institute of Continuum Mechanics
Type: Contribution in non-scientific journal
Journal: Bauingenieur
Volume: 98
Pages: 354-360
No. of pages: 7
ISSN: 0005-6650
Publication date: 2023
Publication status: Published
ASJC Scopus subject areas: Civil and Structural Engineering, Building and Construction
Electronic version(s): https://doi.org/10.37544/0005-6650-2023-11-42 (Access: Closed)
: Details in the research portal "Research@Leibniz University"

BibTeX

@misc{eb1ac9ff2b2e488b89ca5b6d23159700,
title = "Maschinelles Lernen f{\"u}r die numerische Homogenisierung von Beton",
abstract = "Material modeling of concrete using modern numerical methods significantly accelerates the design process of structures. However, for multiscale modeling of such a heterogeneous material, the established homogenization methods are still very computationally intensive, especially for high accuracy requirements. In this paper, we propose a machine learning approach that provides a computationally efficient solution method while delivering a high degree of accuracy. The dataset used for the training and testing process consists of artificial and real microstructural images (input), while the result data (output) are the homogenized stresses of a given representative volume element (RVE). The performance of the model is demonstrated by examples and compared with classical homogenization methods. The developed ML model achieves higher accuracy in determining the homogenized stresses and significantly reduces the computation time.",
author = "Fadi Aldakheel and Michael Haist and Ludger Lohaus and Peter Wriggers",
year = "2023",
doi = "10.37544/0005-6650-2023-11-42",
language = "Deutsch",
volume = "98",
pages = "354--360",
journal = "Bauingenieur",
issn = "0005-6650",
publisher = "Springer-VDI Verlag GmbH und Co. KG",
}