For materials characterization, grain size is one of the essential parameters that has effects on its structural and functional properties. Due to advancements in image acquisition techniques and computer hardware, it is possible to obtain high-quality material microscopy images with greater efficiency. However, analysis of such images often requires data-driven models which then require a large number of annotated images to train the materials-specific models. Such models are largely ineffective when applied to new test materials with different compositions or material characteristics. Therefore, an attempt has been made to develop a model, which can fit a diverse set of materials with optimum accuracy such that the time and effort needed for developing materials-specific models can be drastically reduced as output from such models can provide sparse or even dense annotations.

This research provides insight into the approach to develop a generic grain boundary detection model (GeGra) from material microscopy images. The model is a deep learning model, trained on the dataset containing more than a thousand microscopy images belonging to different materials such as copper, austenite, brass, magnet, hard metal, aluminium etc. and included image acquisition using different microscopy techniques.

The trained model resulted in 70 % Intersection over Union (IoU) score on test dataset when compared against the reference data from subject experts. The result of the GeGra model is also compared with other material-specific models such as for FeNdB permanent magnets and aluminium samples. Further, the output image from the GeGra model can be integrated into external grain size analysis software platforms and has proven to be more than 5 times faster than manual approaches.