Accurate prediction of tensile stresses in repair mortars is vital for the long-term durability of rehabilitated concrete structures. Existing analytical models are based on the material property theory and often struggle to capture the intricate and non-linear behavior exhibited by different mix types used in concrete. To address the limitation of existing models, neural networks were employed as a modelling approach for more robust and versatile predictions. The data used in developing the models was obtained from laboratory experiments. The input variables to the ANN model included: water content, cement, silica fume, superplasticizer, admixture, and age. Three distinct ANN-based models were developed based on: ordinary Portland cement, 10% silica fume as a partial replacement of cement and a combination of the two binder types. These models were evaluated using four performance metrics: coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). When mortars with ordinary Portland cement was used as a binder, the R2, MAE, MAPE, and RMSE were 99.74%, 0.0808, 0.0397, and 0.0138, respectively. For mortars with 10% silica fume, the ANN model predicted restrained shrinkage stresses in mortars with R2, MAE, MAPE, and RMSE values of 99.25%, 0.0090, 0.0731, and 0.3161, respectively. When both binders were used, the R2, MAE, MAPE, and RMSE were 99.77%, 0.0093, 0.0804, and 0.1775, respectively. The application of neural networks for predicting restrained shrinkage stresses in repair mortars outperforms conventional models with enhanced accuracy and reliability. The developed ANN models serve as powerful tools for assessing and optimizing the performance of repair mortars, enabling more efficient and precise design strategies in concrete repair.
Read full abstract