Machine learning-based prediction approaches for the fracture toughness of quasi-brittle materials

Lihua Niu, Nan Li, Junfeng Guan, Chaopeng Xie, Lielie Li, Shanshan Chen,
Machine learning-based prediction approaches for the fracture toughness of quasi-brittle materials,
Engineering Fracture Mechanics,
Volume 327,
2025,
111449,
ISSN 0013-7944,
https://doi.org/10.1016/j.engfracmech.2025.111449.
(https://www.sciencedirect.com/science/article/pii/S0013794425006502)
Abstract: The accurate prediction of fracture toughness in quasi-brittle materials, such as concrete, mortar, and rock, is crucial for ensuring the safety, durability, and long-term stability of civil engineering structures. Fracture toughness is a key parameter that reflects the material’s crack resistance capability and plays a vital role in engineering design. Traditional experimental methods for determining fracture toughness are often time-consuming, expensive, and unable to effectively capture the complex interactions between multiple influencing factors. In recent years, machine learning techniques have become powerful tools for predicting material properties, providing higher predictive accuracy and computational efficiency compared to traditional experimental and empirical approaches. In this study, several machine learning models were developed, trained, and systematically compared to predict the fracture toughness of quasi-brittle materials. The models considered in this study include Multilayer Perceptron (MLP), Support Vector Regression (SVR), Particle Swarm Optimization-based Support Vector Machine (PSO-SVM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). The key predictive features considered include the maximum aggregate size (dmax), water-cement ratio (w/c), volume content of coarse aggregate (Vagg,c), volume content of fine aggregate (Vagg,f), fiber volume fraction (Vf), fiber aspect ratio (γ), height-to-crack ratio (α), and tensile strength (ft). The research involved collecting an extensive dataset from experimental studies, conducting feature preprocessing and selection, and subsequently applying machine learning algorithms to develop predictive models. Cross-validation and multiple performance metrics were employed to evaluate the predictive accuracy and generalization ability of each model. The results indicate that XGBoost model outperform other models in predicting fracture toughness, demonstrating highest prediction accuracy and stronger model stability. Compared to traditional regression analysis methods, machine learning models are more effective at handling non-linear relationships and can fully account for the interactions among multiple parameters. This study highlights the potential of machine learning for predicting fracture toughness in quasi-brittle materials, offering new insights into the integration of intelligent design methods for material optimization and structural safety assessment. By enhancing both the accuracy and efficiency of predictions, the proposed models serve as a valuable tool for optimizing material properties and informing engineering design decisions in civil engineering applications.
Keywords: Quasi-brittle materials; Fracture toughness; Machine learning; Support vector regression; XGBoost