Investigating the Performance of the Attention Mechanism and the Interpretability in the Concrete Strength Prediction Model

To address the limitations of traditional models in capturing complex features for concrete strength prediction, this study proposes a hybrid deep learning approach that integrates multiple attention mechanisms with gated recurrent units (GRU). The methodology employs a multi-scale validation framework, conducting three-dimensional validation across three datasets: the Kaggle standard dataset, the lightweight foam concrete dataset, and the self-compacting concrete dataset. Six attention mechanisms (SE attention, dot-product attention, self-attention, etc.) are comprehensively compared to optimise the GRU network structure. A Newton–Raphson-based optimiser (NRBO) enables hyperparameter adaptive tuning. Experimental results show significant improvements over the baseline GRU model: mean R2 increased by 6.99%, while RMSE and MAE decreased by 38.5% and 37.5%, respectively. SHAP interpretability analysis confirms that attention mechanisms effectively capture key parameters like SP and VMA in the self-compacting concrete dataset. Based on the findings, this study recommends using self-attention for datasets smaller than 200 samples and selecting the higher-accuracy model between self-attention and stacked attention mechanisms for larger datasets.