Abstract Lightweight Strain-Hardening Ultra-High-Performance Concrete (LWSH-UHPC) is a lightweight and durable material that enhances structural efficiency and reduces dead loads in high-rise and bridge structures. This study applies advanced hybrid machine learning (ML) models to predict key properties of LWSH-UHPC, including compressive strength (CS), actual density (DD), theoretical density (TD), material efficiency (ME), production cost, and CO 2 emissions. A total of 248 UHPC datasets were collected from the literature, consisting of 176 conventional and 72 lightweight mixtures. Five hybrid ML models were developed using eight input features: ANFIS-GA, CNN-LSTM, SVM + K-Means, XGBoost + K-Means, and RA-PSO. Among these, ANFIS-GA and CNN-LSTM showed the highest accuracy in predicting compressive strength (80–220 MPa). However, most other models performed less effectively at lower strength levels ( 0.97, R = 0.98), and theoretical density (R 2 = 0.963, MAPE = 1.61 %). Sensitivity analysis identified theoretical density (TD) as the most influential factor affecting compressive strength, with the highest correlation (0.44) and relative importance ( μ = 0.326 ± 0.028). All models achieved high accuracy in predicting production cost and CO 2 emissions (R 2 = 1.000). CNN-LSTM provided the best cost prediction (RMSE = 0.0113, MAE = 0.0091), while XGBoost + K-Means was most accurate for CO 2 emissions (RMSE ≈ 0.0175). The developed ML framework can be integrated into a graphical user interface (GUI) to help engineers quickly estimate mechanical properties, cost, and CO 2 emissions. An open-access data-sharing platform is being developed to promote the use of LWSH-UHPC in bridge and lightweight structural applications.
Aïssa Rezzoug (Thu,) studied this question.