Machine Learning-Based Prediction of Steel–Recycled Aggregate Concrete Bond Stress: A Comparative Study of Artificial Neural Networks and Empirical Models

Abstract

This study investigates the bond behavior between steel reinforcement and recycled aggregate concrete (RAC) using two predictive approaches: empirical formulations and artificial neural networks (ANN). A comprehensive dataset of pull-out experimental test compiled from previous research was utilized to develop four empirical equations, each tailored to a specific range of RAC compressive strength. The equations incorporate key parameters including: concrete compressive strength, the ratio of concrete cover to steel bar diameter, the ratio of steel bar diameter to bond length, the replacement ratio of coarse recycled aggregates, the water-to-cement ratio, and the product of steel bar area and yield strength.

The ANN model was structured with an input layer reflecting the same parameters used in the empirical equations, a hidden layer containing 20 neurons, and an output layer representing bond stress. Training was conducted using the collected dataset, while validation was performed against experimental results from 30 experimental pull-out specimens performed in the present study.

While the empirical models yielded reasonable predictions, the ANN demonstrated superior accuracy and robustness in replicating experimental bond stress values. These findings affirm the potential of ANN as a reliable tool for predicting steel-RAC bond behavior.