EXTENDED ABSTRACT: Introduction: The distribution of material composition is closely linked to the distribution and morphology of microstructural features. The common method for quantitative microstructure characterization involves metallographic microscope observation and manual counting. By combining multi-element coupling with machine learning algorithms, the challenge of non-destructive high-throughput statistical mapping of microstructures can be addressed. This method enables rapid, non-destructive characterization of the type, position, quantity, and dendrite spacing in single crystal high-temperature alloys, while also analyzing the correlation between element distribution and dendritic structures. Method: The production of multi-element coupling descriptors is achieved through the coupling of multidimensional information such as inter element correlation information and element segregation information. A CNN is used to quickly classify primary dendrites, secondary dendrites, and backgrounds by extracting hidden features from the descriptors. Results: The threshold method, BP, and MCF-CNN models were tested respectively, and their classification accuracy rates were 75.11%, 83.16%, and 86.57%, respectively, and the primary dendrite recognition accuracy rates were 78.15%, 95.61%, and 98.10%, respectively. Conclusion: A multi-element coupled MCF-CNN model was built to achieve 86% classiffcation accuracy and 98% primary dendrite recognition accuracy. The number, location, PDAS distribution and component distribution of primary dendrites were quickly quantitatively characterized in multiple dimensions, providing multi-dimensional data support for studying the relationship between organization and performance.

Keywords：Microbeam X-ray; Multi-element coupling; Machine learning; PDAS