A Method Based on Deep Learning for Statistics of Dendrite Spacing in Single Crystal Superalloys

Wan Weihao^1,2,  Li Dongling^1,2,  Wang Haizhou^1,2*,  Zhao Lei^1,2,  Sun DanDan^1,2,  Li jie³

¹ Central Iron and Steel Research Institute；

² NCS Testing Technology Co., Ltd.；

³ University of Science and Technology Beijing

ABSTRACT: Microstructure is an important part of materials. It stores the genes of the material and has a decisive influence on the physical and chemical properties of the material. The dendrite spacing in single crystal superalloys can directly affect the thermomechanical properties and heat treatment process of the alloy. At present, the most direct measurement method is manual measurement,and the Chinese national standard GB/T14999.7-2010 stipulates the method for measuring the spacing . But this method can only obtain the average spacing of dendrites, which Can not fully represent its performance. In order to characterize the dendrite spacing quickly, accurately and comprehensively, a quantitative statistical distribution characterization method for the full field of view of dendrite microstructure in single crystal superalloy was established, as shown in Figure 1.This method consists of two parts: dendrite axis detection and dendrite spacing statistics, which are used to identify primary dendrites in metallographic images and count their dendrite spacings respectively. The Faster R-CNN network, a tobject detection network based on deep learning, is used in the detection process. The detection target is the primary dendrite core. In the statistical process, in addition to calculating the commonly used average spacing, we propose a local multi-directional dendrite spacing calculation method: by finding the nearest dendrite in multiple directions for each dendrite, calculate the spacing in each direction. In this way, the dendrite spacing is refined from the global information of the entire field of view to the local information of each dendrite, which expands the characterizable information of the dendrite spacing. In this work, the dendrite core detection was performed on 4 sample data sets, and the detection results with an average accuracy of 92.685% were obtained, and multiple characteristic parameter information was obtained, such as the local dendrite spacing of the primary dendrite, the average dendrite spacing, and the dendrite Spacing distribution, etc.

Figure1.Identification and statistics process

Keywords: Single crystal superalloy; Dendrite spacing; DeepLearning; Multi-directional statistics