Harnessing AI For Sustainable Construction: Predicting and Optimizing Recycled Aggregate Concrete with Krn-PSO

Abstract

This study presents an AI-driven approach to optimize recycled aggregate concrete (RAC) mixes reinforced with glass fibers (GF). A suite of machine learning (ML) models – including a Keras Recurrent Neural Network (kRN), Multilayer Perceptron (MLP), Back Propagation Neural Network (BPNN), Residual Neural Network (ResNet), and stacked Long Short-Term Memory (LSTM) – was developed to predict the 28-day compressive strength of RAC and identify optimal mix proportions. A comprehensive dataset of RAC mix designs (incorporating up to 100% recycled coarse aggregate replacement and various GF dosages) was compiled from literature and experimental results. Data were preprocessed (normalized and analyzed for correlations), and model hyperparameters were tuned via Bayesian optimization to maximize predictive performance. The best model (kRN) achieved near-perfect accuracy, with an R² ≈ 1.0 on compressive strength prediction. Results confirm an inverse relationship between recycled aggregate content and compressive strength (e.g. 50% RCA caused ~8.9% strength reduction), while the inclusion of GF significantly improved mechanical performance. A small GF addition (0.5% by volume) enhanced compressive strength (e.g. from 35.9 MPa to 37.9 MPa in 0% RCA mixes) and recovered much of the strength lost to RCA, whereas excessive fiber content (>2% GF) led to diminished returns. The optimized ML-guided mix – using 50% recycled aggregate and 0.5–1.0% GF – achieves comparable strength to natural aggregate concrete, with a ~22% boost in tensile strength at 2% GF . These findings demonstrate that ML optimization can effectively balance performance and sustainability in concrete design, enabling high predictive accuracy in compressive strength and guiding the development of greener, fiber-reinforced RAC mixtures.