Machine-learning-based prediction of early-age compressive strength of ultra-high-performance concrete using ultrasonic pulse velocity and rebound hammer

Accurate early-age strength evaluation is essential for ultra-high-performance concrete (UHPC), as construction operations such as formwork removal, prestressing, lifting, and handling depend on timely quality-control decisions. Although ultrasonic pulse velocity (UPV) and rebound hammer testing are practical non-destructive techniques, existing empirical equations often show limited transferability and unsatisfactory accuracy when applied to early-age UHPC. This study proposes a machine-learning-based framework for predicting early-age UHPC compressive strength from non-destructive evaluation data. A unified database compiled from published studies was established, and six machine-learning models were evaluated under two practical input cases using a nested group-aware cross-validation strategy. Case A used age, fiber volume fraction, direct UPV, and rebound number, whereas Case B used only direct UPV and rebound number. The results showed that the best machine-learning models consistently outperformed the existing empirical-equation baselines and provided more reliable prediction under both input conditions. Parameter-importance analysis further showed that fiber volume fraction contributed little relative to age and the two non-destructive evaluation (NDE) indicators. Overall, the proposed approach provides a leakage-resistant prediction framework with practical implementation tools; the simplified explicit equations are intended as transparent approximations when closed-form estimates are preferred.