The prediction and characterization of concrete properties by using the machine learning algorithms: A state-of-the-art review

Concrete strength mainly depends on the hydration between water and cement and how the resulting calcium silicate hydrate (C-S-H) crystals binds the other concrete components together. Traditional empirical formulas can hardly incorporate all relevant factors to accurately predict concrete properties. Over the past two decades, newly developed machine learning algorithms have been applied to practical problems across different fields. These machine learning algorithms offer an alternative approach to predicting concrete properties (e.g., strength and permeability), reconstruct digital twin models, and capture surface defects. This paper reviews the application and improvement of machine learning algorithms in predicting and characterizing the mechanical properties and morphology of concrete, including the traditional machine learning (K-nearest neighbors, support vector machines), artificial neural networks, deep learning models (convolutional neural networks and generative adversarial networks), and physics-informed neural networks. Potential models for the prediction or characterization of a specific property are summarized. This paper will help researchers in the area of concrete materials in selecting and establishing a machine learning model.