AI-Driven Mix Proportioning and Performance Prediction of Low-Cement Ultra-High-Performance Concrete

Abstract

Ultra-high-performance concrete (UHPC) provides exceptional mechanical performance and durability; however, its high Portland
cement content results in increased material consumption and embodied carbon emissions. This study proposes an artificial intelligence (AI)-driven approach for the development and optimization of low-cement UHPC by integrating experimental mix
proportioning, machine learning (ML), explainable AI, and multi-objective optimization. A series of UHPC mixtures incorporating
ground granulated blast-furnace slag (GGBS), silica fume, quartz sand, superplasticizer, and steel fibers was considered to investigate
the influence of binder composition and mixture parameters on fresh and mechanical properties. Multiple ML algorithms, including
multiple linear regression, random forest, support vector regression, artificial neural network, gradient boosting, XGBoost, LightGBM,
and Extra Trees, were evaluated for compressive-strength prediction. Among the models investigated XGBoost gave the best (R^2) of
0.968, RMSE of 3.91 MPa, MAE of 2.87 MPa and MAPE of 2.19%. SHAP explainability showed that curing age, wt/wt ratio, cement
content, silica fume and GGBS were the most important variables influencing compressive strength prediction. The AI-based
optimization produced a low-cement UHPC mixture with about 600 kg/m³ cement, 300 kg/m³ GGBS and 150 kg/m³ silica fume. The
optimized mixture had 28-day compressive strength of approximately 150.6 MPa with 29.4% lower cement consumption compared
to reference UHPC. Estimated embodied CO₂ emissions were reduced by approximately 28.9 % but flowability, tensile strength and
flexural performance were satisfactory. The results show that ML prediction, explainable AI and optimization can provide an effective
data-driven strategy for sustainable UHPC development with low cement consumption and high structural performance.