Bridging Strength Regimes in Recycled-Aggregate Concrete Through Experiments and Machine Learning

Abstract

Recycled concrete aggregate (RCA) is recognized as a means of reducing the demand for virgin aggregates. However, the mortar that often adheres to aggregate particles, the resulting high absorption of RCA, and the heterogeneity of the interfacial transition zones between aggregate and matrix all have the potential to impact the strength of RCA-concrete. Conventional concretes are often classified as either normal- or high-strength based upon their compressive strength. An experimental program that utilized eight different concretes of M30/M80 composition, each with 0, 30, 60, and 100% replacement of natural aggregate with RCA, allowed for the determination of the mean compressive strength of each concrete between 1 and 28 days of curing. A leakage-controlled machine learning approach was developed in conjunction with the experimental analysis to model the relationship between various constituents of the concretes and their resulting strengths. Such variables included the curing age of the concretes, the fraction of RCA within each concrete, the water-to-binder ratio, the amount of binder, Alccofine (fine cement particles) within the concretes, and the dosage of superplasticizer. Each of these variables was used to train an artificial neural network (ANN), a Random Forest, an XGBoost model, and a radial-basis support vector regressor (SVR). The age trajectories of concretes of unseen compositions were used to hold-out samples for cross-validation to assess the models’ accuracies. Results of the experiments indicated that concrete that utilized 30%, 60%, and 100% replacement of natural aggregate with RCA exhibited strength losses of 8.1%, 19.6%, and 23.7% for normal-strength concretes, and strength losses of 12.4%, 29.4%, and 38.8% for high-strength concretes, respectively. The 6-16-8-1 ANN produced the best results (R² = 0.874 ± 0.025, RMSE = 7.12 ± 0.71 MPa, MAE = 5.11 ± 0.48 MPa, and MAPE = 16.17 ± 1.45%), outperforming the SVR, Random Forest, and XGBoost models. The analysis of the models via SHAP diagrams indicated that the curing age of the concretes and the fraction of RCA within the concretes were the two most important variables that impacted the strength of the resulting concretes. Thus, approximately 30% of natural aggregate can be replaced with RCA in the concretes studied. Furthermore, these results indicate that machine learning approaches can be used as means of predicting the strength of concretes containing RCA, but with consideration of the leakage within the datasets and the generally limited generalizability of results based upon small sample sizes and averaging of results among samples.