Machine learning-driven prediction of mechanical properties of lightweight concrete based on experimental data.

Lightweight concrete (LWC) is increasingly used in structural and non-structural applications due to its ability to reduce self-weight while maintaining adequate mechanical performance. However, predicting the mechanical behavior of EPS-based lightweight concrete remains challenging because of the complex and nonlinear interactions between mix composition, density, curing age, and aggregate-matrix characteristics. This study proposes an integrated experimental and machine learning-based framework to predict the mechanical properties of lightweight concrete incorporating expanded polystyrene (EPS) particles (Addipor 55) as a partial volumetric replacement of natural coarse aggregate. An experimental program was conducted using EPS replacement levels ranging from 0 to 500 L/m³, combined with a fixed silica fume content of 60 kg/m³ and a polycarboxylate-based superplasticizer at a constant water-cement ratio of 0.35. Compressive strength, splitting tensile strength, and density were measured at curing ages of 3, 7, and 28 days. The results showed a systematic reduction in density from 2380 kg/m³ for the control mix to 1720 kg/m³ at the highest EPS content. The 28-day compressive strength decreased from 41.2 MPa to 28.6 MPa, while the splitting tensile strength declined from 4.30 MPa to 2.65 MPa. To enhance model robustness, the experimental dataset was expanded using a physically constrained data augmentation approach based on experimentally observed trends. An Artificial Neural Network (ANN) model was then developed and validated using independent data subsets, demonstrating excellent predictive performance with coefficients of determination (R²) of approximately 0.998 for both compressive and splitting tensile strengths. The proposed ANN model, supported by a graphical user interface (GUI), provides a reliable and efficient tool for predicting the mechanical performance of EPS-based lightweight concrete and optimizing mix design with reduced experimental effort.