Towards rapid fatigue assessment of UHPC for wind turbine towers: A sparse-data-driven prediction framework via DeepONet with latent vector adaptation

Abstract

Rapid assessment of high-cycle fatigue degradation in ultra-high-performance concrete (UHPC) is constrained by the long duration of full-lifecycle fatigue testing. This study develops an uncertainty-aware Bayesian transfer learning framework (U-BTL) that combines Deep Operator Networks with specimen-specific latent adaptation to reconstruct long-term stiffness-degradation trajectories from sparse early-cycle observations. During Stage I, operator learning is supported by measured source trajectories, low-weight Weibull-IDW augmentation restricted to the experimentally supported source domain, an exact initial-stiffness boundary K ( 0 ) = 1 , an irreversible-degradation constraint, and a terminal-stiffness surrogate. During Stage II, the pretrained operator is frozen and a specimen-specific latent vector is calibrated probabilistically from a fixed early-cycle observation window. The posterior uncertainty is subsequently propagated to long-term stiffness and threshold-crossing predictions. The evaluation distinguishes interpolation validation from fully held-out out-of-distribution conditions involving an unseen fiber configuration and an unseen stress level. The results demonstrate that Bayesian latent adaptation can reduce specimen-specific prediction bias while providing probabilistic uncertainty bounds for future degradation trajectories. The present validation is restricted to constant-amplitude laboratory UHPC specimens; therefore, the proposed framework is interpreted as a constitutive-level fatigue-prognosis kernel whose future integration into wind-turbine structural monitoring requires additional variable-amplitude, structural-scale, and field validation.