A sound understanding of the basic principles which dictate dendritic solidification of Ni-base SXs is crucial, because the dendritic microstructure causes a large scale microstructural heterogeneity which can affect the resistance against creep and high temperature fatigue loading [e.g. 1]. Different types of detrimental defects, e.g. freckles, slivers and elongated stray grains, which accompany the directional solidification of SX, have been studied extensively [2-4]. Recently crystal mosaicity has received new attention [5-7]. Crystal mosaicity in SXs is associated with slightly misoriented dendrites which are separated by interdendritic regions containing small angle grain boundaries. In order to understand the cast microstructure, one must study interactions between individual dendrites during solidification. As dendrite populations evolve, competitive growth processes can be observed, where individual dendrites are influenced by each other. To be able to understand how competitive dendrite growth governs crystal mosaicity three-dimensional experimental and theoretical approaches are required. The associated data collection and curation is tedious and time consuming. Therefore, in the present work, an automated procedure was developed consisting of (1) deep learning object detection, (2) image registration and (3) the design of an appropriate morphological data base structure.

The raw data was obtained from metallographic cross sections taken from directionally solidified cylindrical SX specimens. Cross sections were taken by slicing the cylinder in 1 mm distances. Optical micrographs (16 per cross section combined into montages) of the whole cross section containing several thousand dendrites were recorded with a resolution of 5000x5000 pixels. We present different approaches of how to handle the large amount of raw data and compare the object detection performances of multiple network architectures and various hyper parameters. Finally, we condense the information by transforming it into a sparse data base of geometric representations reflecting all relevant geometric relations between dendrites. This novel approach allows us to perform further automatic data extraction and sets us up for developing the method further using big data mining approaches.