Microstructure-informed deep learning model for accurate prediction of multiple concrete properties

Predicting multiple properties of concrete using empirical models has become increasingly challenging due to the complexity of modern concrete formulations and the nonlinear behavior of their constituents. This study introduces a sequential model that integrates mix proportions with microstructural information of concrete. The model addresses the limitations of small datasets and the inherent variability in concrete's raw materials and production processes. A novel dataset comprising concrete mix proportions, 56,160 scanning electron microscope images, and their corresponding macroscopic properties was constructed for training and validation. We developed a sequential model integrating a Swin Transformer (Swin-T) with a Back Propagation Neural Network (BPNN), achieving superior accuracy in predicting compressive strength and permeability. Comprehensive evaluations using SHAP and GradCAM reveal the critical role of hydration products in these predictions, underscoring the enhanced interpretability and efficacy of our approach. This work advocates for the integration of microstructural insights to improve the reliability and precision of concrete assessments.