Time-resolved prediction and optimization of sustainable concrete strength using machine learning and genetic algorithm

Abstract

Eco-friendly concrete is a sustainable construction material designed to reduce environmental impact by incorporating recycled materials and minimizing carbon emissions. However, traditional empirical methods often fail to accurately predict its performance due to the complex interactions among novel additives such as glass fiber and marble dust. This study presents an integrated experimental and machine learning framework to predict and optimise concrete’s compressive, flexural, and split tensile strengths over 7, 14, 28, and 56-day curing periods. Advanced models, Artificial Neural Networks (ANN), Support Vector Regression (SVR), Random Forests (RF), Extreme Gradient Boosting (XGBoost) and hybrid CNN-LSTM (Convolution Neural Networks and Long Short Term Memory) were evaluated. Among these, the hybrid CNN-LST demonstrated superior performance, achieving R² values of 0.999, 0.999, and 0.999 for compressive, flexural, and split tensile strengths, respectively, with a minimum RMSE of 0.0095 for compressive strength prediction. Feature importance analysis revealed curing time as the most influential variable, while the sensitivity analysis suggested optimal strength to be maximum at approximately 8–10 kg of marble dust and 15–21 kg of glass fiber. A multi-objective Genetic Algorithm (GA) and NSGA—II (Non -dominated sorting algorithm) were used to optimize the mix design, yielding predicted 56-day strengths of 37.24 MPa (compressive), 4.27 MPa (flexural), and 3.42 MPa (split tensile). Monte Carlo simulations were used to assess the uncertainty and enhance robustness. The proposed framework significantly reduces the experimental workload while offering a cost-effective, scalable strategy for developing sustainable high-performance concrete using industrial waste.