Deep learning-based prediction of compressive strength of mortar containing silica flour: Integrating various temperature exposure scenarios

Various structures such as buildings and bridges continue to suffer from increased permeability and compressive strength loss at higher temperatures, commonly referred to as cement strength retrogression. To avoid this condition, silica flour (SF) has been proposed to enforce the formations of stable, crystalline structures. Building upon experimental research, this research employs deep learning (DL) models (deep interest network (DIN) Transformer, convolutional neural network (CNN), neural oblivious decision ensemble (NODE), and feature tokenizer (FT) Transformer) to evaluate the compressive strength of mortar incorporating SF at different temperatures. SHapley Additive exPlanations (SHAP) analysis is further employed to develop interpretable and transparent insights based on best-testing-performing model, while 5-fold cross-validation is employed to generate precise estimates of model performances. The FT-Transformer model achieves the highest overall predictive performance, with an R² score of 0.996 over the training data set, whereas CNN achieves the highest testing prediction (R² = 0.9753). The SHAP analysis highlights the impacts of age, cement, water, silica flour, and temperature on model performance. It is also demonstrated that CNN residual analysis can successfully produce excellent model prediction, suggesting that the residuals are uncorrelated with consistent variance. Moreover, the 5-fold cross-validation results reveal that fold 1 achieves excellent generalisation ability among all the folds, with the lowest gap of −0.0064, while the remaining 4 folds are predominantly characterised by strong generalisation ability, with positive gaps not exceeding 0.02. Consequently, the DL models used in this study provide a new paradigm for effective forecast of the compressive strength of mortar incorporating SF at various temperatures.