Myung Soo Kang, Changgil Lee, Yun‐Kyu An
Abstract Thermal stress estimation in steel structures is critical for in-situ integrity assessment because temperature variations under constrained conditions can induce substantial internal stresses. Despite its importance, estimating thermal stress of in-situ structures remains challenging because the coupling of temperature- and stress-induced effects on structural responses obscures the true stress state. Therefore, this study proposes a guided wave-based thermal stress estimation that integrates a virtual digital model (VDM) with a deep learning-driven virtual emulator (VE). The VDM generates synthetic guided wave responses over a temperature range under both constrained and unconstrained boundary conditions. To mitigate the discrepancies between the simulated and measured signals, the framework exploits temperature-induced variations in the guided wave features rather than relying on absolute waveforms. Accordingly, the simulated responses are converted to temperature–time-of-flight graphs ( TTGs ): thermal stress-dependent graph ( TT G F ) and temperature-dependent graph ( TT G R ). The VE is trained to infer TT G R from TT G F , reconstructing the temperature-only guided wave feature from measurements under constrained conditions, decoupling temperature and stress effects. Two Gaussian process regression models are trained to map TT G F to the coefficient of thermal expansion ( α ) and TT G R to Young’s modulus, which are key properties governing thermal stress. In practice, TTG F ′ is obtained from a constrained structure, and the corresponding TTG R ′ is generated using the trained VE. These TTG F ′ and