IntroductionThe precise characterization of microstructure in cement-slag blended pastes is pivotal for understanding reaction mechanisms and optimizing material performance. Traditional methods relying on single-model analysis face limitations in distinguishing phases with similar grey value ranges and integrating multi element mappings. This study proposes an intelligent Backscattered electron and Energy-Dispersive X-ray Spectroscopy (BSE-EDS) image analysis methodology, integrating standardized pre-processing procedures, Potential of Heat-diffusion for Affinity-based Trajectory Embedding and Gaussian Mixture Model (PHATE-GMM) based phase identification model, and Glue-based Multiview visualization, to achieve multimodal characterization of cement-slag pastes. The framework bridges the gap between the BSE image and EDS element mappings with multimodal features in cementitious systems, enabling comprehensive phase identification, element migration analysis, chemical composition analysis, phase evolution tracking, and microstructure quantitative characterization.MethodsCement-slag paste (30% slag replacement ratio, water-to-binder ratio 0.4) were prepared, cured (0, 7 d and 14 d), and examined by Scanning electron microscope (SEM) after epoxy impregnating and polishing. The BSE image, qualitative element mappings, and quantified element mappings were acquired with accelerating voltage of 15 kV, working distance of 11 mm. A BSE-EDS data preprocessing pipeline was developed: (1) Guided filtering enhanced EDS element mapping resolution using BSE image as guided image; (2) Generating phase masks excluding pore and epoxy interference; (3) Super-pixel segmentation on phase region (10 000 units) to integrates grey value information from BSE images and element information from quantified element mappings. The PHATE-GMM model combined PHATE dimensionality reduction algorithm for 2D visualization with GMM for clustering PHATE-derived attributes, achieving automated and interpretable phase identification. Glue platform enabled synchronized multi-view interactive analysis (element scatterplots, phase maps, PHATE plots, element histograms, density plots) to validate phase identification results and explore phase relationships.Results and discussionPHATE-GMM model achieved great consistency with visual inspection-labeled phases. In raw materials, distinct clusters corresponded to cement clinker, slag and gypsum. The C₃A/C₄AF phases are frequently embedded within the C₂S/C₃S phases, which results in partial super-pixel regions containing both phases during super-pixel segmentation. Consequently, interconnected clusters can be observed in PHATE plot. For hydrated pastes, rim phases (e.g., C₂S/C₃S-rim, slag-rim, CH-rim) emerged as transitional zones between critical phases and C-S-H. The spatial distribution and element composition analyses of both slag-rim and slag demonstrate that the slag-rim exhibits a lightly higher Mg/Ca ratio compared to slag, indicating a distinct enrichment of Mg ions in the rim region. Local analysis mode via Glue facilitates element exploration in un-clustered cementitious systems, particularly for complex or novel binders. Global analysis mode via Glue, suitable for well-characterized cementitious systems. PHATE-derived element density maps further enable visualization of ion migration pathways. The phase network topology diagram achieves tri-dimensional synergistic representation through spatial positioning, radial scale, and connection line thickness parameters, while preserving the original relational information between phases shown in PHATE plots and simultaneously incorporating phase content features. The BSE-EDS image intelligent analysis methodology established in this study enables efficient and accurate determination of pore and phase in cement-slag binders, thereby enabling multiple downstream applications including elemental composition analysis of phases, investigation of evolutionary patterns, and quantitative characterization of microstructure.Conclusions1) The proposed pipeline for generating BSE-EDS datasets demonstrates strong generalizability and practicality, effectively mitigating the impact of pore interference and excessive interaction volume on phase identification, providing a viable data management solution for multimodal feature extraction and downstream analysis in cementitious systems.2) The developed PHATE-GMM phase identification model exhibits high interpretability and intelligence, enabling comprehensible automated phase identification. The visualization structure of PHATE plot offers novel analytical insights for scientifically describing reaction mechanisms and microstructure evolution in cementitious systems.3) The Glue-based multimodal feature analysis technique facilitates synchronous and interactive exploration of chemical and spatial modalities, allowing for expert-guided optimization of automated phase clustering results.4) The intelligent BSE-EDS image analysis methodology enables comprehensive downstream applications, which fully exploits the analytical potential of BSE-EDS for microstructural characterization, realizing the technical vision of "single independent micro-characterization - multimodal feature fusion analysis".