Interfacial Bonding Properties of Natural Fiber Concrete Using Vision Transformer

Chunmei Yao, Xiangru Dong

Article
2026 / Volume 9 / Pages 1333-1357
Received 28 August 2025; Accepted 13 November 2025; Published 29 April 2026
https://doi.org/10.31881/TLR.2026.1333

Abstract
This study addresses the challenge of predicting the interfacial bond strength in natural fiber concrete by analyzing microscopic images. Traditional methods struggle to establish a reliable link between fine-scale interface characteristics and overall performance, especially when multiple types of fibers are involved. We develop a novel approach based on a Vision Transformer model to analyze scanning electron and optical micrographs of interfaces involving jute, sisal, coconut coir, flax, and hemp fibers. The image analysis framework is designed to be robust to variations among different fibers, enhancing its ability to generalize. Experimental results show that the model achieves accurate bond strength predictions on the test set, with a mean absolute error of 0.095, a root mean square error of 0.140, and a coefficient of determination of 0.975. The model maintains low prediction errors even for fiber types not included during training, demonstrating strong generalization. Analysis of the model’s focus confirms that it identifies physically meaningful features at the fiber-matrix interface, with coconut coir fiber showing the least interfacial activity, correlating with its measured strength. This work provides an intelligent and interpretable tool for studying and optimizing natural fiber concrete.

Keywords
natural fiber concrete, vision transformer, bond strength prediction, domain adaptation, microscopic images