Practical prediction models for the deformation-related properties of recycled aggregate concrete

The mechanical behaviour of recycled aggregate concrete (RAC) is more complex than that of natural aggregate concrete (NAC) due to the inclusion of residual mortar and interfacial transition zones (ITZs) adjacent to it. This complexity leads to a trade-off in the prediction equations between simplicity and high accuracy, both of which are important for structural designers. Traditional regression-based equations and machine-learning-based prediction models have their limitations: the former highly depends on the parameter ranges of the data used for regression and lacks generalizability, while the latter may be unreliable when training data is not enough (a few hundred data for the behavior of RAC is far below the thousands generally required for machine learning). To address this gap, using the deformation-related properties of RAC as an example, this study proposes a method to simplify the theoretical formulas by replacing variables with representative values from statistical analysis of 2088 collected aggregate properties and 324 concrete mix proportions. The models developed with this method combine advantages of simplicity, generalizability, and high accuracy. Validation results showed that the proposed models achieved mean absolute percentage error (MAPE) values of 9.5 % for 489 elastic modulus samples, 9.4 % for 144 shrinkage samples, 10.1 % for 61 creep samples, and 10.6 % for 178 peak strain samples.