Macro and mesoscopic mechanical behavior of concrete with actual aggregate segmented by hybrid Transformers and convolutional neural networks

The precise establishment of mesoscale models for concrete is crucial for understanding its mechanical behavior through numerical simulations. This study presents a deep learning-based numerical framework aimed at investigating both macro- and mesoscopic mechanical behavior of concrete, focusing on...

Full description

Saved in:
Bibliographic Details
Main Authors: Dong Wang, Junxing Zheng, Jichen Zhong, Lin Gao, Shuling Huang, Jiajia Zheng
Format: Article
Language:English
Published: Elsevier 2025-07-01
Series:Case Studies in Construction Materials
Subjects:
Online Access:http://www.sciencedirect.com/science/article/pii/S2214509525002128
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:The precise establishment of mesoscale models for concrete is crucial for understanding its mechanical behavior through numerical simulations. This study presents a deep learning-based numerical framework aimed at investigating both macro- and mesoscopic mechanical behavior of concrete, focusing on the actual aggregate morphology and distribution in 2D cross-sections. A dual encoder concrete aggregate segmentation network (DECAS-Net) based on CNN and Transformer is developed to segment aggregates and construct a refined mesoscale model compatible with the discrete element method (DEM). A multi-branch parallel dilated convolutional feature fusion module is designed to effectively integrate multi-scale features from the different encoders. Uniaxial compression tests are conducted to simulate the effects of actual aggregate shape and distribution on the macro- and mesoscopic mechanical behavior of concrete. Comparative analyses are performed utilizing models with circular aggregates and randomly distributed irregular aggregates, providing further insights into the influence of actual aggregate morphology and distribution on concrete mechanical behavior. The results demonstrate that the proposed DECAS-Net method provides accurate foundational data for establishing refined 2D mesoscale models, and highlights the significant impact of aggregate morphology and distribution on the mechanical behavior of concrete.
ISSN:2214-5095