Strength and durability prediction and multi-objective mix design optimization of recycled aggregate concrete based on machine learning

The extensive application of recycled aggregate concrete (RAC) in the construction industry has gained significant attention currently due to its potential environmental and economic benefits. However, numerous time-consuming experiments are required to evaluate the properties of RAC. Furthermore, multiple conflicting objectives need to be considered for RAC mix design, such as high mechanical strength, high durability, low carbon footprint, and cost-efficiency. Thus, this study employs nine machine learning models to predict the compressive strength (CS), carbonation depth (CD), and passing electric charge (PEC) of RAC after hyper-parameter optimization, utilizing datasets of 965, 698, and 231 samples, respectively. LightGBM achieved the highest performance for CS (R²=0.9523, RMSE=2.7163) and CD (R²=0.9286, RMSE=1.8723), while XGBoost excelled in predicting PEC (R²=0.8985, RMSE=514.3255). Generally, the predicted values of selected models and the actual values were well fitted, with an overall control at around 20%. The selected machine learning models are superior to conventional quantitative models. By combining the objectives of carbon footprint and cost, the Adaptive Geometry Estimation based MOEA II (AGE-MOEA-II) algorithm was used for multi-objective optimization. A hybrid combination of Analytic Hierarchy Process (AHP) and Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) methods was employed to obtain the optimal solution in two design scenarios. This proposed framework can be applied to various engineering applications, which could improve the efficiency in the mix design optimization and promote the use of RAC in engineering.