Leveraging machine learning to evaluate the effect of raw materials on the compressive strength of ultra-high-performance concrete

• Machine learning models showed strong predictive accuracy for UHPC compressive strength. • XGB outperformed RF, GB, and GPR with the highest R-value and the lowest RMSE. • Curing age, silica fume, and fiber content positively impact UHPC strength. Ultra-High-Performance Concrete (UHPC) is distinguished by its exceptional mechanical strength and durability, making it a preferred material for high-performance structural applications. However, the high cement content required for achieving these properties significantly contributes to carbon emissions, posing environmental concerns. To enhance the sustainability of UHPC, it is imperative to develop strategies for reducing cement consumption while maintaining its superior performance. The shift toward sustainable construction demands innovative strategies to balance environmental and performance goals. Machine learning accelerates this transition by delivering data-driven insights and optimizing material design, fostering eco-friendly construction solutions. This study presents an approach to forecasting the compressive strength (CS) of UHPC, ranging from 270 kg/m³ to 750 kg/m³, by employing advanced machine learning techniques. The research utilizes stepwise linear regression (SLR), random forest (RF), gradient boosting (GB), extreme gradient boosting (XGB), and Gaussian process regression (GPR) algorithms, developed based on 357 experimental results sourced from recent studies. The impact of 12 influential features on CS was evaluated to optimize the performance of the proposed models. Among the algorithms, XGB outperformed the others with an R² of 90.1 % and a lower RMSE of 11.52 MPa, surpassing RF (88.7 %), GB (89.2 %), and GPR (86.1 %). Feature importance analysis revealed that steel fiber content, curing age, and silica fume content are the most significant factors affecting CS, while quartz powder and limestone powder showed minimal influence. Additionally, Partial Dependence Plots were utilized to quantitatively assess the contribution of each input variable to the CS of UHPC. These findings offer valuable guidance for design engineers and construction professionals, enabling data-driven optimization of UHPC mixes to achieve both sustainability and high mechanical performance.