Statistical and Machine-Learning Approaches for Predicting Concrete Compressive Strength: A Review of Methods, Predictors, and Practical Performance

الملخص

Concrete compressive strength is the primary criterion used to judge the quality, safety, and durability of concrete structures, yet its conventional 28-day determination delays quality decisions on site. Over the past two decades a substantial body of research has used statistical regression and, increasingly, machine-learning methods to predict compressive strength directly from mixture proportions and curing age. This review synthesizes that literature across three broad families of numerical approach: SPSS-based multiple linear regression on single-laboratory datasets, classical correlation and regression studies that established the field's core predictors, and more recent work using larger, pooled, or heterogeneous databases together with nonlinear and machine-learning models. Across more than thirty reviewed studies, the water-cement ratio and cement content emerge as the most consistent predictors of strength, while curing age contributes a positive but nonlinear, diminishing-returns effect. A clear and recurring pattern links reported model accuracy to dataset scale and homogeneity: small, single-source datasets tend to yield very high coefficients of determination (R² > 0.90), whereas large, pooled, multi-source databases yield more moderate values (R² between 0.50 to 0.65), which likely better represent the predictive accuracy achievable in general engineering practice. The review identifies persistent gaps around dataset heterogeneity, limited treatment of supplementary cementitious materials and admixtures, sparse interaction-term modelling, and inconsistent independent validation practice, and proposes a research agenda addressing these gaps. The findings are intended to help practicing engineers and researchers to select an appropriate predictive approach and to calibrate their expectations of accuracy according to the scale and diversity of the underlying dataset.

These findings aim to guide practicing engineers and researchers in selecting appropriate predictive approaches, thereby calibrating accuracy expectations based on data diversity