A Critical Review of Machine Learning Approaches for Prediction of Mechanical Properties of Concrete

Abstract

Compressive strength and the related mechanical properties of concrete govern how a structure is designed, checked and maintained, yet they are still established mainly through destructive tests that take days to weeks. Machine learning (ML) offers a way of estimating these properties from mixture proportions, curing history and, increasingly, non-destructive test data. This review covers the prediction of compressive strength, flexural and splitting tensile strength and the modulus of elasticity across the recent studies reviewed. Thirty studies are compiled in a single table, grouped by the property predicted and covering material type, dataset size, inputs, model and reported accuracy. Four recent developments are examined: the second generation of boosted libraries, automated model selection, pre-trained tabular models that return a predictive distribution rather than a point estimate, and the first genuine reinforcement learning applications to mixture design. Compressive strength dominates the compiled evidence, while the modulus of elasticity and the tensile and flexural properties rest on far fewer studies and far smaller datasets, so their reported accuracies deserve more caution. The central finding is that the evaluation protocol now limits confidence in reported accuracies more than the choice of algorithm does. Duplicate records left in place inflate R² from roughly 0.60–0.71 to 0.81–0.87, and the leading boosted ensembles differ by less than the spread across data splits.