A Feature Informed Lifetime Prediction Framework for Thermal Barrier Coatings via Multimodal Fusion Characterization

ABSTRACT Thermal barrier coatings (TBCs) play a critical role in ensuring the long‐term reliability of hot‐section components in aero‐engines and gas turbines operated under extreme service environments. Accurate assessment of their service lifetime remains challenging, particularly under complex degradation conditions involving CMAS (CaO–MgO–Al 2 O 3 –SiO 2 ) attack, where the evolution of microstructural damage is strongly coupled with thermal and chemical loading. Conventional lifetime prediction approaches primarily rely on empirical experience or destructive testing, limiting their ability to capture the intrinsic relationship between damage mechanisms and lifetime degradation. In this study, a feature‐informed lifetime prediction framework is developed by integrating multimodal characterization data to evaluate the residual lifetime of TBCs subjected to CMAS attack. Service‐related parameters, including ceramic layer thickness, columnar grain spacing, service temperature, and CMAS compositional ratios, are employed to describe the external loading conditions and internal structural states governing damage evolution. In parallel, microstructural features extracted from post‐service microscopic images are introduced to characterize the actual damage response of the coating, capturing key degradation phenomena such as crack propagation, columnar grain deterioration, and CMAS infiltration. Deep microstructural descriptors are further obtained using a convolutional neural network to quantitatively represent complex damage patterns induced by coupled thermal and chemical effects. Service parameter features and image‐derived microstructural features are treated independently at the feature level and subsequently fused within a regression‐based modeling framework, enabling a quantitative correlation between microstructural damage evolution and residual lifetime degradation. The proposed multimodal fusion approach shows significant improvement in prediction accuracy and robustness compared with models based on single‐feature modalities, as evidenced by an RMSE of 4.69 and an R 2 of 0.98. These results highlight the potential of multimodal feature integration for reliable lifetime assessment of TBCs under complex CMAS service conditions.