A deep learning fusion model for in-situ grain size analysis of nickel-based superalloys

In the material characterization of superalloys, widely accepted methods for grain size detection include the comparative method, the area method, and the intercept method. However, all of these methods rely on manual implementation. These methods are relatively time-consuming and subjectively biased. Under this prerequisite, it is very challenging to design and implement an efficient method to assist in determining the grain size of superalloys materials. In this study, a deep learning method for extracting grains from scanning electron microscope images with in-situ analysis was proposed. The method was able to derive the variation rule of its grain size in mechanical experiments. We took visual geometry group 16 as the backbone network and fused the convolutional networks for biomedical image segmentation to get the grain boundary segmentation and extraction model. This model was trained using the scanning electron microscope images obtained from the in-situ tensile experiment. The proposed model showed higher performance with the dice similarity coefficient of 81.8 % when validated on unforeseen images than other baseline models. Moreover, the model was capable of accurately segmenting and extracting twins that exhibited no discernible grain boundaries and were distinguished solely by brightness contrast. The areas of grains were also calculated and the rule of change of grain size was summarized on the results with in-situ analysis. This method can be applicable to a variety of grain size detection methods and expanded for different needs.