Assessment of AutoML frameworks for predicting compressive and flexural strength of recycled aggregate concrete

Abstract

The use of recycled aggregate concrete (RAC) is crucial for promoting sustainable construction practices by mitigating the environmental impact associated with the extraction of natural aggregates (NA) and reducing ${\mathrm{CO}}_2$ emissions. This study aims to evaluate the performance of five automated machine learning (AutoML) frameworks — H2O, AutoKeras, FLAML, TPOT, and AutoGluon — in predicting the properties of RAC. The dataset comprises 638 samples with 13 variables, including compressive strength (CS) and flexural strength (FS). The results indicate that AutoKeras, based on deep learning, performed poorly due to the small dataset size and high dimensionality, which are not ideal for deep learning models. In contrast, FLAML and H2O demonstrated superior performance, with FLAML achieving the highest $R^2$ (0.780) and lowest RMSE (6.928) for CS predictions. The Tukey test confirmed significant differences between AutoKeras and the other models, while AutoGluon, FLAML, H2O, and TPOT showed comparable effectiveness. This study highlights the importance of selecting appropriate AutoML models for accurate and reliable RAC property predictions, contributing to the reduction of ${\mathrm{CO}}_2$ emissions, conservation of natural resources, and promotion of a circular economy in the construction sector.