Concrete compressive strength is a predictor of structural performance, durability, and safety. The Best strength prediction is required for the best mix design, maximum structural reliability, and minimum construction cost. Concrete strength is influenced by several mix constituents; their interactions are nonlinear and limit the application of empirical and regression-based methods. To compensate for these problems, this study employs high-performance Machine Learning (ML) models in conjunction with a data-driven optimization strategy to well predict concrete compressive strength. The data for concrete composition and age were split into training data of 80% and test data of 20%, and Standard Scaler normalization was conducted to enhance model robustness and learning efficiency. ML models were developed with the Arithmetic Optimization Algorithm (AOA) optimization for the best predictive performance. The novelty of this research lies in the application of AOA optimization for ensemble ML models with faster convergence, improved accuracy, and enhanced generalization compared to traditional optimizers such as Genetic Algorithm (GA). Among the AOA-optimized models, CatBoost was the best, predicting compressive strength compared to other AOA-optimized ML models within ± 20% of experimental values for 89.81% of the test data with R2 of 0.983, and MAE of 2.635, also superior to its GA-optimized counterpart, achieving <InlineEquation ID="IEq1"> <EquationSource Format="TEX">${a}_{20}\_index$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msub> <mi>a</mi> <mn>20</mn> </msub> <mi>_</mi> <mi>i</mi> <mi>n</mi> <mi>d</mi> <mi>e</mi> <mi>x</mi> </mrow> </math></EquationSource> </InlineEquation> of 78.64%, R2 of 0.8798. Additionally, permutation importance and Explainable Artificial Intelligence (XAI) results were utilized to connect model predictions with hydration chemistry and mix design principles, offering engineering actionable recommendations. This combination of techniques is novel in the field of construction material informatics. The results show that ML-based forecasts can minimize experimental effort, optimize resource utilization, and facilitate robust, economical, and sustainable concrete mix design in actual building settings.

Integrating interpretability, using hybrid data-driven ensemble learning, and optimization framework for concrete strength prediction

期刊论文
发表日期：

2025年12月13日

查看全文

He Jia,
Wang Qin

摘要

Concrete compressive strength is a predictor of structural performance, durability, and safety. The Best strength prediction is required for the best mix design, maximum structural reliability, and minimum construction cost. Concrete strength is influenced by several mix constituents; their interactions are nonlinear and limit the application of empirical and regression-based methods. To compensate for these problems, this study employs high-performance Machine Learning (ML) models in conjunction with a data-driven optimization strategy to well predict concrete compressive strength. The data for concrete composition and age were split into training data of 80% and test data of 20%, and Standard Scaler normalization was conducted to enhance model robustness and learning efficiency. ML models were developed with the Arithmetic Optimization Algorithm (AOA) optimization for the best predictive performance. The novelty of this research lies in the application of AOA optimization for ensemble ML models with faster convergence, improved accuracy, and enhanced generalization compared to traditional optimizers such as Genetic Algorithm (GA). Among the AOA-optimized models, CatBoost was the best, predicting compressive strength compared to other AOA-optimized ML models within ± 20% of experimental values for 89.81% of the test data with R² of 0.983, and MAE of 2.635, also superior to its GA-optimized counterpart, achieving ${a}_{20}\_index$ $a_{20}_i n d e x$ of 78.64%, R² of 0.8798. Additionally, permutation importance and Explainable Artificial Intelligence (XAI) results were utilized to connect model predictions with hydration chemistry and mix design principles, offering engineering actionable recommendations. This combination of techniques is novel in the field of construction material informatics. The results show that ML-based forecasts can minimize experimental effort, optimize resource utilization, and facilitate robust, economical, and sustainable concrete mix design in actual building settings.