Machine learning (ML) has become an increasingly important tool in concrete engineering which has significantly altered the method of prediction and optimization of concrete properties, enabling more efficient, accurate, and sustainable processes. However, the inherent variability of concrete is a significant challenge to the generalization and performance of ML models. This study is a review that explores the effect of the variability of concrete material on the reliability and accuracy of predictions by ML. To explain the influence of these sources of variability on mechanical and durability related behaviors, the paper groups the sources of variability into four major groups, namely composition, microstructure, curing conditions, and environmental factors. A broad range of machine learning paradigms—including supervised learning, unsupervised learning, reinforcement learning (RL), and hybrid physics-informed approaches—is examined with respect to their robustness against data heterogeneity and distributional shifts. The weaknesses and advantages of the two types of algorithms are highlighted with regard to forecasting fresh and hardened concrete properties and the optimization of the mix design. Based on this synthesis, the review identifies key unresolved challenges, including the lack of standardized multi-source datasets, limited transferability of models across experimental settings, and insufficient reporting of preprocessing and normalization practices.
Bahmani et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: