A physics-informed machine learning framework for accelerated discovery of single-phase B2 multi-principal element intermetallics

Abstract

Single-phase ordered body-centered cubic or B2 multi-principal element intermetallics (MPEIs) have garnered significant attention due to their exceptional mechanical and functional properties. However, their discovery in complex compositional spaces is challenging due to the lack of high-dimensional phase diagrams and the inefficiency of traditional trial-and-error methods. In this study, we developed a physics-informed machine learning (ML) framework that integrates a conditional variational autoencoder (CVAE) with an artificial neural network (ANN). This approach effectively addresses the challenges of data limitation and imbalance, enabling the high-throughput generation of B2 MPEIs. Using this framework, we successfully identified a wide range of B2 complex alloys, spanning quaternary to senary systems, with superior mechanical performance. This work not only demonstrates a significant advancement in the discovery of B2 MPEIs but also provides an accelerated pathway for their design and development.