Ultrasonic defect detection in a concrete slab assisted by physics-informed neural networks

Traditional nondestructive testing (NDT) methods face challenges to accurately assess concrete owing to its naturally inhomogeneous nature that complicates spatial characterization of material properties. To address these limitations, this work considers physics-informed neural networks (PINNs) interpreting contactless ultrasonic scan data to enhance defect detection capabilities in concrete. PINNs integrate physics laws through mathematical governing equations into artificial neural network models to overcome limitations of purely data-driven analysis approaches. The study utilizes experimental data collected from a large-scale concrete slab containing inclusion, cold joints with cracks, and surface fire damage and from a homogeneous PMMA slab (as a reference). The PINN results are used to create space-dependent property maps based on the extracted coefficient of the governing wave equation using a simple time-domain wavefield data set. The results demonstrate that PINNs effectively predict space-dependent wave velocities. This approach facilitates accurate material property characterization and defect identification. The proposed PINN models achieved a P-wave velocity prediction error of 0.34 % for the PMMA slab and identified areal extent of defects in the concrete slab with errors of 1 % for pristine areas and 2.1 % for inclusion areas. Sub-wavelength-sized cracks around the inclusion areas were detected from the predicted wave velocity map. These findings suggest that PINNs offer a promising approach for improving the accuracy and efficiency of defect detection in concrete structures with superior spatial resolution provided by other conventional ultrasonic imaging approaches.