Mechanical property prediction of superalloys with microporosity defects using a multi-source deep learning framework

This study proposed a novel multi-source deep learning framework for predicting the tensile properties of superalloys containing microporosity defects by simultaneously incorporating both microstructure and defect features. A comprehensive multi-source dataset was constructed using multi-source microstructure and microporosity defect images obtained from tensile specimens extracted from cast plates with diverse microstructural and defect characteristics. The target mechanical properties include ultimate tensile strength (UTS), yield strength (YS), and elongation (EL). Compared to models trained using only microstructure or defect images, the proposed multi-source framework achieved superior prediction accuracy, with R2 values exceeding 0.93 for all three properties. In addition, the mean absolute error (MAE) decreased with an increasing number of microstructural image channels, indicating that the incorporation of multi-source microstructural features significantly enhances model performance. Furthermore, an explainable AI methodology was applied to reveal the underlying mechanisms by which microstructural and defect features govern tensile behavior. This framework presents a data-driven methodology for uncovering microstructure-defect-property relationships, providing a potential pathway for accurate mechanical property prediction of defect-containing superalloys. Its modular architecture can also be readily applied to other alloys.