Machine Learning–Based Prediction of Compressive Strength in Sustainable Concrete Incorporating Industrial By-Products: A Critical Review

Abstract

The partial replacement of Portland cement with industrial by-products such as fly ash, ground granulated blast furnace slag, silica fume and calcined clay is now a mainstream decarbonisation strategy in concrete technology. It also complicates strength prediction: by-product blends behave non-linearly, mature at different rates, and vary in composition from source to source, so the water–cement ratio rules embedded in conventional mix design lose much of their explanatory power. Machine learning (ML) has been widely proposed as the remedy. This paper reviews the development of ML-based compressive strength prediction for concretes containing industrial by-products, tracing the field from early artificial neural network models through ensemble and gradient boosting methods to hybrid and metaheuristic-optimised architectures, and to the recent turn toward interpretable models. The review argues that reported accuracy has plateaued while methodological quality has not kept pace. Four recurring weaknesses are identified: heavy dependence on a small number of reused public datasets, description of by-products by mass dosage rather than chemical or physical characteristics, near-universal absence of external validation on independently produced mixes, and reporting conventions that favour headline coefficients of determination over error distributions and uncertainty. Interpretability methods have improved transparency but are frequently used to confirm expectations rather than to test them. The paper concludes with a research agenda emphasising composition-aware feature sets, external and temporal validation, uncertainty quantification, and integration of prediction with multi-objective mix optimisation so that data-driven models contribute to sustainability outcomes and not only to prediction scores.