MACHINE LEARNING-BASED PREDICTION OF CONCRETE STRUCTURAL PERFORMANCE: A CRITICAL REVIEW OF MODELS, APPLICATIONS, AND CHALLENGES

Abstract

Concrete remains the world's most widely used construction material, yet predicting its structural performance under various conditions presents significant challenges due to complex nonlinear relationships between material composition, environmental factors, and mechanical properties. Traditional experimental and empirical methods for assessing concrete strength, durability, and structural integrity are time-consuming, costly, and labor-intensive. Recent advances in machine learning (ML) and deep learning (DL) have emerged as powerful alternatives for modeling concrete structural performance with unprecedented accuracy and efficiency. This review synthesizes developments in ML applications for concrete structural prediction, examining neural network architectures, ensemble methods, hybrid optimization frameworks, and explainable artificial intelligence techniques. The paper critically analyzes 60+ peer-reviewed studies covering compressive strength prediction, tensile strength estimation, fracture energy prediction, durability modeling, crack detection, seismic damage assessment, and structural health monitoring. Key findings demonstrate that ensemble methods—particularly XGBoost, CatBoost, and Random Forest—achieve prediction accuracies exceeding R² = 0.96 across diverse concrete formulations. However, significant challenges persist, including limited field-based datasets, poor generalization across material systems, data imbalance issues, and insufficient model explainability. This review identifies critical research gaps and recommends future directions toward physics-informed machine learning, multi-task learning frameworks, integration of non-destructive testing data, and development of open-access benchmark datasets to advance the field toward practical deployment in civil engineering design and infrastructure management.