Deep learning accelerates the development of Ni-based single crystal superalloys: A physical-constrained neural network for creep rupture life prediction

Abstract

Creep rupture life is a key target for designing Ni-based single crystal (SX) superalloys. Machine learning (ML) has the potential to accelerate alloy design by accurately predicting the creep rupture life of potential superalloys. In this work, a physical-constrained neural network named Superalloy Transformer-based Network for Creep (SaTNC) is established, and the representation of superalloy is automatically learned based on deep representation learning. As core physical bias in the SaTNC model, the information associated with interaction among alloying elements achieved by Transformer-based neural network is introduced to constrain the representation of the superalloy. The SaTNC model exhibits competitive generalization performance and the results have verified the high capability of the SaTNC model to capture the sensitivities of creep rupture life on alloy composition and service conditions. Moreover, Transformer-based deep representation for superalloy with both effectiveness and interpretability is obtained. The material insights including the interactions among alloying elements and their combined effects on the creep rupture life of superalloy are extracted from this deep representation. Finally, with the aid of the SaTNC model, a potential alloy composition space with low density and excellent creep resistance at high temperatures is identified, which will be experimentally verified in the future.