Deep Learning Architectures for Concrete Compressive Strength Prediction: A State‑of‑the‑Art Review of CNN, ANN, and Hybrid Models

Abstract

Structural safety, optimization of materials, and sustainable construction practice depend on the prediction of concrete compressive strength. Traditional methods of testing use the laboratory method, which is time‑consuming, expensive, and destructive. Recent progress in deep learning has made it possible to predict the compressive strength in an accurate, rapid, and non‑destructive way by modeling nonlinear complex relationships between the constituents of concrete, curing conditions, and mechanical performance. This review is a systematic review of the deep learning architectures that have been applied to predict the concrete compressive strength with state‑of‑the‑art, such as Convolutional Neural Networks (CNNs), Artiϐicial Neural Networks (ANNs), Deep Neural Networks (DNNs), Long Short‑Term Memory (LSTM) networks, Gated Recurrent Units (GRU), Transformer‑based models, and hybrid architectures (CNN‑LSTM, CNN‑GRU, and ensemble stacking). It has been shown in the literature that higher hybrid and ensemble models allow the high predictive performance to be achieved, with the value of R² often exceeding 0.95, with the best possible models having an R² of 0.99 when using controlled datasets. Both metaheuristic optimization algorithms (e.g., PSO, GA, ACO, TLBO) and Bayesian hyperparameter tuning would greatly increase the model generalization and robustness. Moreover, interpretable artificial intelligence tools, such as SHAP and sensitivity analysis, have enhanced interpretability, and cement content, curing age, and water‑cement ratio are confirmed to be the most significant predictors of strength. Applications have been spread over the spe‑ cialized materials like ultra‑high‑performance concrete (UHPC), geopolymer concrete, recycled aggregate concrete, self‑compacting concrete, and waste‑based sustainable concretes. However, the issues of data standardization, cross‑laboratory generalization, and model transparency persist in spite of impressive advances. The future research is to be directed at physics‑informed neural networks, the multi‑objective optimization that considers the metrics of environmental impact, real‑time edge deployment, and the standardized benchmark datasets. In general, methods using deep learning as its core technology can be discussed as a revolutionary development in intelligent concrete design and sustainable construction engineering