What type of study is this?

October 9, 2025Open Access

Prediction of compressive strength of high-performance concrete based on multiple machine learning models

KXKai XiaoSouthern University of Science and Technology HZHongjian ZhangZhongshan Hospital of Xiamen University SWSijia WeiXinjiang University

Key Points

HHO-XGB model provides the highest predictive accuracy for compressive strength in high-performance concrete.
R² increases from 0.91 to 0.95 when comparing the Decision Tree model to the HHO-XGB model, indicating significant improvement.
The study uses a dataset of 1,030 samples and evaluates model performance with R², MAE, and RMSE metrics.
Validation through k-fold cross-validation ensures the robustness of the machine learning models used in the study.

Abstract

High-performance concrete (HPC) exhibits excellent comprehensive performance and is widely applied in tunnel engineering, large-span bridges, and special engineering projects. With the advancement of technology, HPC is moving towards green and sustainable development by incorporating industrial solid waste as a supplementary cementitious material. This study constructs machine learning models (individual and ensemble learners) to predict the compressive strength of HPC. The database employed in this study includes eight parameters (including cement, blast furnace slag, fly ash, water, superplasticizer, coarse aggregate, fine aggregate, age), with a total of 1,030 data samples. This study evaluates the performance of the constructed models using the coefficient of determination ( R 2 ), mean absolute error ( MAE ), and root mean square error ( RMSE ), and validates the models using k-fold cross-validation (k = 10). The results indicate that the Decision Tree (DT) model has the best predictive performance among individual learners, while the Harris Hawks Optimization-XGBoost (HHO-XGB) model has the best performance among ensemble learners. The ensemble learning further improves the predictive performance of individual learners: compared with the best individual learner (DT), R 2 increases from 0.91 to 0.94 (Random Forest (RF)) and 0.95 (HHO-XGB); MAE decreases from 2.72 MPa to 2.69 MPa (RF) and 2.51 MPa (HHO-XGB); RMSE decreases from 5.01 MPa to 4.01 MPa (RF) and 3.57 MPa (HHO-XGB), respectively. In addition, the constructed models have been validated for robustness through k-fold cross-validation. The superior predictive accuracy of the HHO-XGB model can provide a more reliable basis for optimizing mix designs, thereby enhancing structural safety and reducing material cost overruns in critical applications like tunnel linings and marine structures.

KI fragen

Bookmark

View Full Paper