Physics-grounded virtual mock-up and simulation-to-real transfer framework for AI-assisted PAUT of nuclear power plant carbon steel pipe welds

This study presents a physics-grounded virtual mock-up and simulation-to-real framework for AI-assisted phased array ultrasonic testing (PAUT) of carbon-steel pipe welds in nuclear power plants. The work addresses two obstacles to digital inspection deployment: the scarcity of defect-containing data and the cost and scalability limits of physical mock-ups. The target application is representative secondary-side 6-in and 10-in class carbon-steel butt welds in Korean pressurized water reactors, especially OPR1000 and APR1400 units, rather than primary reactor coolant loop piping. A virtual mock-up was established to reproduce the weld geometry, transducer, wedge, and scan plan of an inspection specimen. Its suitability for transfer learning was assessed by comparing simulated and measured S-scan images using three complementary similarity metrics: VGG16 for global structural correspondence, FSIM for local structural contrast, and Gabor similarity for texture consistency. The results indicate strong global agreement (VGG16 > 93%) but a micro-texture discrepancy (FSIM <80%), highlighting the difficulty of reproducing stochastic grain-noise and diffraction features. Based on the validated configuration, 11,816 simulated S-scan images were generated and labeled for seven defect classes. Among the evaluated detectors, YOLOX provided the most favorable compromise between geometric adaptability and inference efficiency. When applied to independently fabricated weld specimens acquired with the same PAUT system, the simulation-trained model detected 11 of 12 defects (91.7%) and yielded root-mean-square deviations from nominal design values of approximately 10 mm in length and 3.5 mm in height. These findings support validated virtual mock-ups as conservative screening assets, while indicating that improved noise modeling and future domain-adaptation strategies may be needed for low-SNR volumetric flaws.