A methodology for stress-strain behavior characterization and mixture optimization of recycled aggregate concrete based on machine learning

Abstract

The characterization of stress-strain behaviors of recycled aggregate concrete (RAC) excluding physical properties of recycled aggregates (RAs) may result in an inaccurate prediction of mechanical responses in practical applications. In this study, a data-driven model using a refined long short-term memory (LSTM) network is established based on the Bayesian optimization algorithm, with the motivation to accurately predict the uniaxial compressive stress-strain behaviors of RAC, including the stress-strain relation, elastic modulus, peak stress, and the peak strain. Training and testing of the proposed model require the integration of the mixture content and the fundamental physical properties of RAs with the stress-strain relation of ordinary concrete featured by prominent sequential attributes. Upon a dataset containing 100 experimental samples from independent studies, covering a wide range of RA substitution rates, the superior prediction capability of the proposed LSTM network is demonstrated in comparison with the analytical results of three empirical mechanics-driven models. Finally, the trained LSTM network is further employed to optimize the mixture for RAC using the Bayesian optimization technique innovatively, to achieve a balance between the mechanical performance and requirement to the quality of RAs. This study constitutes a beneficial addition to the understanding of the RAs-dependent stress-strain behaviors of RAC from the perspective of the data paradigm and provides a practical tool for the optimal design of RAC materials and structures.