Machine learning prediction of compressive and flexural strength of plastic aggregate concrete using DGAN-augmented data

The increasing accumulation of plastic waste has encouraged its reuse as a partial aggregate replacement in concrete, offering a sustainable alternative to conventional construction materials. However, the complex relationships between plastic content, mixture proportions, and concrete strength, together with limited experimental data, make reliable strength prediction challenging. This study develops a machine-learning framework for predicting the 28-day compressive strength (CS) and flexural strength (FS) of HDPE plastic aggregate concrete. The study used 260 original compressive-strength (CS) and 178 original flexural-strength (FS) observations, which were augmented using DGAN to address the limited data availability. Machine-learning models were evaluated using an internal train-test and cross-validation framework; however, independent external or leave-one-study-out validation was not performed. Five machine-learning models; Decision Tree, ANN, MLP, CatBoost Regressor, and Stacked Ensemble were evaluated using multiple performance metrics, while SHAP and PDP were used to interpret model predictions. The Stacked Ensemble achieved the best CS prediction (R 2 = 0.990, RMSE = 1.061 MPa, MAE = 0.732 MPa), whereas CatBoost provided the highest FS prediction (R 2 = 0.946, RMSE = 0.240 MPa, MAE = 0.095 MPa). Water-cement ratio, cement dosage, and plastic-related parameters were identified as important contributors to predicted strength. The novelty of this study lies in integrating DGAN-based data augmentation, comparative machine learning, and model interpretation for HDPE-containing plastic aggregate concrete. The proposed framework demonstrates the potential of data-driven methods for strength prediction under limited-data conditions and provides a basis for further supplementary experimental verification and model development.