Physics assisted machine learning methods for predicting the cracking performance of recycled aggregate concrete

Abstract The poor cracking performance has become the obstacle for the application of recycled aggregate concrete (RAC), but its prediction is hard to implement. In this study, the physics assisted machine learning methods are adopted to predict the cracking performance of RAC. With the assistance of physics, 9 features are effectively selected as the inputting variables, the splitting tensile strength is selected as characterization parameter of cracking. The CART, SVR, Adaboost and Random Forests algorithm are used to construct the predictive models, the Firefly algorithm is used to search the optimum hyperparameters. By comparing the predicted value and experimental data during training and testing procedure, the Adaboost model is proved to be the excellent model for predicting the tensile strength of RAC. Combined with the physical mechanism, the important analysis proves that the contents of recycled aggregates, aggregate size and water contents are three most influential factors for the predictive models, and these factors should be carefully considered during designing the mixture of RAC. Moreover, the reliability of predictive models is verified by physical experiments.