Abstract
Predicting the long-term deformation of structural materials under extreme conditions remains a grand challenge in materials science, especially for refractory alloys, where high-temperature creep limits performance and service life. Here, a physics-informed digital twin framework is developed that integrates a viscoplastic self-consistent (VPSC) model, real-time high-temperature creep experiments, and a calibration neural network to predict and elucidate the creep behavior of Mo-14Re alloys. The digital twin accurately reproduces creep curves across 1000-1200 °C and 60-150 MPa, achieving <5% deviation from experiments. Crucially, the learned parameter trajectories uncover a previously unrecognized mechanism: Re solute atoms are dragged by gliding dislocations ("solute-drag" effect), leading to rhenium segregation at grain boundaries and compromised creep strength. This is corroborated by post-mortem TEM and molecular dynamics simulations. Furthermore, the model-guided strategy reveals that nanoscale La(2)O(3) precipitates can pin dislocations and suppress Re segregation, significantly improving creep resistance. This work advances the mechanistic understanding of refractory alloy creep and demonstrates a transferable AI-enabled digital twin approach for materials design under extreme environments.