Accurate prediction of concrete compressive strength plays a significant role in quality control during construction. Previous predictive models largely focused on the influence of initial mix proportions

but neglected the impact of aggregate gradation

derived features

and its interpretability. This study develops a stacking-based deep ensemble model for predicting compressive strength

which holistically considers these two factors to improve predictive accuracy and interpretability. This novel model uses three widely used ensemble learning algorithms and a Convolutional Neural Network (CNN) as heterogeneous base learners to leverage diversity and heterogeneity among these algorithms. To improve the tree-based models that are usually too sensitive to hyperparameters and limited by their capacity for high-dimensional feature extraction

we integrate CNN with a channel attention mechanism

thereby enhancing its feature representation capability. A Multi-Layer Perceptron (MLP) incorporating an attention mechanism is adopted as a robust meta-learner to mitigate overfitting risks. Leveraging the SHAP (Shapley Additive explanation) framework

we examine systematically the critical features of concrete strength prediction and their interactive effects. Experimental results show our new model

through considering aggregate gradation and derived features comprehensively

achieves a 27.53% improvement in the accuracy of compressive strength predictions. Using a SHAP analysis

we have identified the dominant drivers of the model－water-to-binder ratio

water content

fly-ash-to-water ratio

cement content

and the mass fraction of aggregates in the size range of 31.5-40 mm. This study improves predictive accuracy and sheds light on the understanding of core parameters governing concrete strength through interpretable analysis

helping intelligent concrete management.