Dataset scope and experimental record diversity
The framework relies on 1,204 experimental records capturing diverse material designs and temperature exposure situations. These records span varied mixture compositions, fiber types, curing conditions, and temperature levels, providing a foundation for model training and evaluation.
Eight-model comparison and ensemble superiority
The study evaluates eight machine-learning models for compressive strength prediction. Ensemble approaches consistently outperform individual learners across all metrics. Both gradient boosting and extreme gradient boosting achieve an R² of 0.903 on unseen data, indicating strong predictive accuracy.
Error metrics and model selection
Gradient boosting produces the minimal root mean square error of 6.67 MPa. Extreme gradient boosting achieves the lowest mean absolute error of 3.97 MPa and the lowest mean absolute percentage error of 10.60%. These figures highlight the complementary strengths of each ensemble method.
Interpretability through SHAP and partial dependence
Shapley additive explanations and partial dependence analysis reveal temperature as the primary factor influencing compressive strength. Fine aggregate, superplasticizer, cement content, water-to-binder ratio, and water content follow as secondary contributors. The analysis shows that strength deterioration with rising temperature varies depending on the specific mix design.
Python graphical interface for prediction
The best-performing models are wrapped in a Python-based graphical user interface, enabling expedited strength prediction without requiring direct code interaction. This accessibility supports practical deployment in design workflows.
Methodology for temperature-dependent design
The study offers a methodology for evaluating temperature-dependent compressive strength. The data-driven approach facilitates the design of fiber-reinforced high-performance concrete across varied exposure conditions.