Machine learning to estimate the tensile strength of cementitious composites

authored by: Minjin Cai, Yuhang Liu, Hehua Zhu, Timon Rabczuk, Shuwei Zhou, Xiaoying Zhuang
Abstract: The application of machine learning models for predicting the tensile performance of Engineered Cementitious Composites (ECCs) holds great promise for advancing both research and engineering practice. However, there remains a lack of systematic evaluation of the predictive accuracy of different machine learning approaches for this task. This study addresses this gap by conducting a comprehensive analysis of seven widely used models: Support Vector Regression, Random Forest, Neural Network, Decision Tree, Gradient Boosting, Elastic Net, and Bayesian Ridge. Using a dataset comprising 129 samples, the models were rigorously evaluated through training and experimental validation. The results indicate that Random Forest and Gradient Boosting models consistently demonstrate strong generalization capabilities and high predictive accuracy, making them well-suited for practical applications. In contrast, Neural Network and Decision Tree models, despite achieving high accuracy on the training data, tended to suffer from overfitting, leading to weaker performance on unseen data. It is important to note that the performance limitations of the Neural Network are largely attributable to the relatively small dataset used in this study. Neural networks typically require larger datasets to fully realize their generalization potential; therefore, the observed results do not imply inherent limitations of the method itself. Elastic Net and Bayesian Ridge models showed the lowest predictive accuracy, suggesting the need for further optimization. Support Vector Regression yielded moderate performance, with balanced predictive accuracy and generalization ability. This study provides a comparative framework to guide practitioners in selecting appropriate models for specific applications, ensuring improved reliability and accuracy in predicting ECC performance.
Organisation(s): Computational Science and Simulation Technology
External Organisation(s): Tongji University
Bauhaus-Universität Weimar
Type: Article
Journal: Structures
Volume: 80
No. of pages: 18
ISSN: 2352-0124
Publication date: 10.2025
Publication status: Published
Peer reviewed: Yes
ASJC Scopus subject areas: Architecture, Civil and Structural Engineering, Building and Construction, Safety, Risk, Reliability and Quality
Electronic version(s): https://doi.org/10.1016/j.istruc.2025.110032 (Access: Closed)
: Details in the research portal "Research@Leibniz University"

BibTeX

@article{53b5887bb4024a97a2414351592c7083,
title = "Machine learning to estimate the tensile strength of cementitious composites",
abstract = "The application of machine learning models for predicting the tensile performance of Engineered Cementitious Composites (ECCs) holds great promise for advancing both research and engineering practice. However, there remains a lack of systematic evaluation of the predictive accuracy of different machine learning approaches for this task. This study addresses this gap by conducting a comprehensive analysis of seven widely used models: Support Vector Regression, Random Forest, Neural Network, Decision Tree, Gradient Boosting, Elastic Net, and Bayesian Ridge. Using a dataset comprising 129 samples, the models were rigorously evaluated through training and experimental validation. The results indicate that Random Forest and Gradient Boosting models consistently demonstrate strong generalization capabilities and high predictive accuracy, making them well-suited for practical applications. In contrast, Neural Network and Decision Tree models, despite achieving high accuracy on the training data, tended to suffer from overfitting, leading to weaker performance on unseen data. It is important to note that the performance limitations of the Neural Network are largely attributable to the relatively small dataset used in this study. Neural networks typically require larger datasets to fully realize their generalization potential; therefore, the observed results do not imply inherent limitations of the method itself. Elastic Net and Bayesian Ridge models showed the lowest predictive accuracy, suggesting the need for further optimization. Support Vector Regression yielded moderate performance, with balanced predictive accuracy and generalization ability. This study provides a comparative framework to guide practitioners in selecting appropriate models for specific applications, ensuring improved reliability and accuracy in predicting ECC performance.",
keywords = "Accuracy evaluation, ECC, Machine learning models, Performance prediction",
author = "Minjin Cai and Yuhang Liu and Hehua Zhu and Timon Rabczuk and Shuwei Zhou and Xiaoying Zhuang",
note = "Publisher Copyright: {\textcopyright} 2025",
year = "2025",
month = oct,
doi = "10.1016/j.istruc.2025.110032",
language = "English",
volume = "80",
journal = "Structures",
issn = "2352-0124",
publisher = "Elsevier Ltd.",
}

Back