Rapid estimation of γ' solvus temperature for composition design of Ni-based superalloy via physics-informed generative artificial intelligence

Abstract

The exceptional high-temperature mechanical properties of Ni-based superalloys are mainly stemmed from the L1₂ γ' phase, therefore it is crucial to discover Ni-based superalloys with high γ' solvus temperatures. Utilizing generative artificial intelligence, we have developed a framework to swiftly evaluate the γ' solvus temperature and tailor Ni-based superalloys, accelerating the process of discovering Ni-based superalloys. Physics-informed artificial neural network emerged as the optimal choice for reverse engineering, outperforming other models with an R² score of 0.917 and a mean absolute error of 15 K. In the reverse design process, 20,000 virtual alloy samples were generated based on divide-and-conquer variational autoencoder which divides the dataset into distinct clusters by K-means algorithm provides a structured representation of the alloy composition space, thereby facilitating a more nuanced understanding of its inherent complexities. In a specific alloy design example, 563 samples were identified through screening based on criteria like γ' solvus temperature, composition deviation index, price, and density. Thermodynamic calculations were used to further screen Ni-based superalloys with exceptional high-temperature properties. The showcase of BA alloy discovery through generative artificial intelligence demonstrates the potential of our research to steer the creation of novel compositions for Ni-based superalloys with outstanding high-temperature properties.