Deep Learning-Based High-Throughput Quantitative Characterization of Multi-Scale γ′ Phase and Correlation with High-Temperature Creep Behavior in Polycrystalline Superalloys

The distribution of multi-scale γ' phases in polycrystalline superalloys is one of the key factors affecting their high-temperature mechanical properties, therefore the characterization of the size, quantity, and distribution of multi-scale γ' phases is crucial. GH4198 superalloy used in third-generation aircraft engine turbine disks was chosen as a research object and a quantitative statistical distribution characterization method for multi-scale γ' phases based on image deep learning was established. Multi-scale γ' phase particles were extracted from the substrate surface by in-situ electrolytic etching method and rapid acquisition of secondary electron images of multi-scale γ' phase particles was achieved through high-throughput scanning electron microscopy. The Real-ESRGAN model was introduced to perform super-resolution processing on secondary electron images of the γ' phase, significantly improving the pixel resolution of nanoscale γ' phase particles. Significant improvement in training set data was achieved through data augmentation methods such as random cropping, rotation, and flipping. Based on the U-Net network structure, a γ' phase intelligent segmentation model was obtained through 100 iterations of training, achieving rapid acquisition and statistical distribution analysis of the position, quantity, and size of multi-scale γ' phases in the entire field of view. The quantitative analysis accuracy of this method exceeds 90%, which is higher than that of traditional methods. At the same time, the analytical results of the nano-precipitate phases also have good consistency with the characterization results of X-ray small angle scattering method and transmission electron microscopy. Furthermore, high-temperature creep tests were performed on specimens from different parts of the GH4198 turbine disk and the influences of γ' phase size and distribution on alloy creep behavior were investigated.