Compositional distribution prediction of heterogeneous materials using graph neural network

Abstract Composition testing is a critical step in steel making industry. In this paper a novel graph neural network (GNN) based approach is presented for predicting the compositional distribution of heterogeneous materials like steel slag. The method uniquely integrates image data and spectral data to address the inefficiency of traditional single-point or exhaustive sampling techniques. By converting images into graph structured patches and fusing them with spectral embedding, the model selectively prioritizes regions with high morphological variability such as oxidation boundaries and pores, while suppressing redundant homogeneous areas. The method achieves measurement precision comparable to full-coverage scans by processing only 30% of data required by conventional methods. The method has been validated on 3000 steel slag samples with multi-oxide detection (e.g. SiO₂, CaO). The framework bridges the gap between laboratory grade accuracy and field deployable efficiency, offering a scalable solution for slag recycling and quality control in steel production.