Prediction of Ultra-High-Performance Concrete (UHPC) Compressive Strength in Structural Applications Using Machine Learning: A Comprehensive Evaluation via Multiple Random Seeds and Hyperparameter Optimization

Ultra-high-performance concrete (UHPC) is increasingly being used in academic and commercial applications due to its superior mechanical properties, high durability, and long service life. To ensure that UHPC can be reliably used in structural applications, it is important to predict its compressive strength in advance. The complex and nonlinear effects of different mix components and their proportions on compressive strength make strength prediction a significant engineering challenge. Reliable compressive strength prediction is crucial for supporting trial mix selection during construction processes, particularly for novel UHPC mixtures. Although there have been studies on predicting UHPC strength using machine learning (ML), the sensitivity of model performance to random seeds and its statistical reliability following hyperparameter optimization have been examined only to a limited extent. This study compares 19 regression models using a dataset of 810 records representing 219 distinct mix designs with constituent quantities, age, and temperature as inputs. A mixture-based repeated split procedure was employed, reserving 20% of the mixtures for testing. Hyperparameter optimization via Optuna utilized a grouped five-fold cross-validation method on the remaining mixtures, incorporating preprocessing steps into the model at each fold. The LightGBM model achieved the lowest mean held-out root mean square error (RMSE) of 10.8838 ± 0.8861 MPa and a coefficient of determination of 0.9187 ± 0.0155. Baseline models relying solely on age, or on age combined with temperature, yielded RMSE values of 28.768 MPa and 24.672 MPa, respectively, indicating that mix composition provides valuable predictive information. Dependence-corrected tests incorporating the Holm correction identified eight significant differences between the selected reference model and the others across 18 comparisons; however, the results showed that this model did not differ statistically from the top-performing ensemble models. Reference model selection and testing were performed using the same results, and a single dataset was reused across the partitions; therefore, the absence of a statistically significant difference does not imply that the models are equivalent. The findings support the preliminary screening of candidate UHPC mixtures based on compressive strength and their prioritization for experimental studies. This approach can contribute to early-stage material evaluation; however, the predicted strength values do not directly represent structural capacity, durability, or construction safety, nor do they replace engineering design checks.