Defect-parameterized physics-informed neural network for forward and inverse modeling of laser ultrasonic wavefield

Abstract

Accurate and efficient non-destructive evaluation (NDE) techniques are essential for ensuring the structural integrity of metallic components, where surface defects, as one of the most common types, can severely compromise fatigue life and mechanical performance. Laser ultrasonics provides a non-contact and high-fidelity approach for inspecting metallic components through wavefield-based analysis. However, reconstructing complete ultrasonic fields and identifying defects from limited measurements remain challenging problems. This study proposes a defect-parameterized physics-informed neural network (DP-PINN) for the forward and inverse modeling of laser ultrasonic wavefield, with the objective of characterizing sub-millimeter surface defects in metallic components. The proposed framework embeds defect-related parameters into the governing elastodynamic equations to reconstruct the full wavefield and estimate the wave velocity field, thereby revealing defect characteristics including location and size. To comprehensively assess the method's performance, four different defect cases are simulated, incorporating different defect characteristics. Furthermore, six practical scenarios are analyzed based on different levels of prior knowledge about material properties and data sparsity. Results demonstrate that defect characterization and full wavefield reconstruction can be achieved with limited measurement data of 0.42 MB. The proposed method maintains consistent detectability across varying defect cases and yields a mean Intersection over Union (IoU) of 0.387, indicating quantitative accuracy.