Machine learning model for predicting the influence of crystallographic orientation on thermomechanical fatigue of Ni-base superalloys

Abstract

Ni-base superalloys used in gas turbine blades are often directionally-solidified (DS) or single-crystal (SX) to minimize grain boundaries and enhance creep resistance. However, their anisotropic elastic properties cause significant variability in mechanical response depending on the loading direction, particularly for components with complex geometries. Predicting thermomechanical fatigue (TMF) and creep-fatigue life across different crystallographic orientations is challenging due to numerous influencing parameters. This work extends a previously developed probabilistic physics-guided neural network (PPgNN), originally limited to $〈001〉$ orientations, by incorporating a normalized elastic modulus as a physics-informed input to account for crystallographic orientation. The model further includes transverse loading in DS alloys. For the first time, a general PPgNN framework predicts both cycles to failure and life variance, integrating critical parameters such as strain range, temperature extremes, dwell durations, cyclic frequency, TMF phasing, and crystallographic orientation. Unlike existing empirical models, which address only a subset of these factors, the proposed approach unifies all key parameters within a single framework. Physics-informed feature engineering, combined with a novel loss function and constrained neural network architecture, enables robust generalization from sparse datasets, providing reliable life predictions across all orientations and test conditions.