Alkali-activated binder concrete strength prediction using hybrid-deep learning along with shapely additive explanations and uncertainty analysis

Alkali-activated binders have been widely used as they are low-carbon materials and eco-friendly to environment. Compressive strength (CS) is the most crucial parameter for designing the alkali-activated binder (AAB) concrete. Hence, the accurate prediction of CS becomes the most important when considering the issues like time saving and cost benefits analytical point of view. This study adopted the conventional deep learning (CDL) and hybrid deep learning (HDL) models to enhance the prediction of CS through analyzing the experimental CS data collected from previous research articles. Three CDL models such as CNN, GRU and LSTM, and their hybrid HDL forms like CNN-GRU, CNN-LSTM and LSTM-GRU were employed to predict the CS property. The models were run with 2060 data points including eight input parameters of AAB concrete specimens, where the 70% data points were used for training and 30% for testing phases. It was observed that HDL models showed the better prediction performance than the CDL models, where the hybrid CNN-LSTM model presents the significant accurate prediction of strength property, with R²= 0.997, RMSE = 1.45 and R²= 0.995, RMSE = 2.24 for training and testing phases, respectively. Moreover, SHAP analysis of CNN-LSTM model exhibits that age of concrete specimens is the main dominant parameter for strength development and cement contents is the second highest contributing issue to strength increment of alkali-activated concrete. These results might be applied in the field of building construction sectors for optimum design of alkali-activated concrete specimens for better strength achievement.