Predicting high-cycle fatigue strength of precipitation-hardened Nickel-Based superalloys from transfer learning

Abstract

Addressing the challenges faced in high-cycle fatigue experimentation through information technology is a desired advancement in the research of Ni-based superalloys. In the present study, a transfer-learning (TR) convolutional neural network (CNN) model is established to break through the limitation of generalization performance in the case of a very small dataset by utilizing material prior information for fatigue strength (FS) prediction. A two-tiered TR framework was implemented, training CNNs on a large tensile dataset to develop source-trained models. This model was then fine-tuned on a smaller, specialized fatigue dataset, resulting in robust models for predicting FSs. It was found that prior to model transfer, the average training and testing accuracies on the large tensile dataset exceeded 95% and 85%, respectively. Following the transfer, the TRCNN model achieved an average testing accuracy of 92.0% on the small fatigue dataset, significantly outperforming the non-transfer CNN model, which recorded an accuracy of only 58.1%. The trained FS prediction models are employed in conjunction with optimization algorithms to forecast compositions of IN718 that exhibit enhanced FSs within a predefined feature space. A new IN718 alloy composition with a room temperature fatigue strength of 636 MPa was identified, surpassing the existing value by more than 25%. Further inquiry showed increasing the Fe content while decreasing the Ti and Nb content can enhance the fatigue strength of precipitation-hardened nickel-based superalloys at room temperature. It also suggests that extending the solution and aging times contribute to the enhancement of the alloy’s fatigue resistance.