A data-augmented ensemble learning framework for predicting ultra-high-performance concrete compressive strength

Ultra-high-performance concrete (UHPC) attracted attention owing to its potential to enhance structural efficiency and service life while reducing material consumption. However, the complex interactions among constituent materials and the limited availability of experimental data make accurate strength prediction and mix design optimization challenging. To address these challenges, this study integrates advanced data augmentation and machine learning to improve the accuracy and reliability of UHPC strength predictions. A comprehensive dataset of 104 compressive strength tests was compiled as the modeling foundation. A boundary-preserving data augmentation method expanded the dataset to 500 samples while maintaining original statistical properties. Five ensemble machine learning algorithms were employed, with extreme gradient boosting (XGB) achieving the highest accuracy. Model interpretability analyses were conducted to reveal key factors influencing strength. This framework, combining data acquisition, augmentation, modeling and interpretability, offers new insights and practical guidance for understanding UHPC behavior, optimizing mix designs, and supporting its engineering applications.