Sparse wavefield reconstruction and defect indication in low-SNR laser ultrasonics using physics-informed neural...

Full-optical laser ultrasonics (LOU) provides broadband, non-contact wavefield measurements but is practically limited by low signal-to-noise ratios (SNR) and long scanning times. In these conditions, conventional interpolation and purely data-driven networks often fail to recover physically consistent fields. This study experimentally investigates a Kirchhoff-Love-plate-constrained physics-informed neural network (KL-PINN) framework for sparse wavefield reconstruction and residual-based defect indication under simultaneous low-SNR and sparse-sampling conditions in full-optical LOU. By embedding a Kirchhoff-Love thin-plate equation into the learning process, the KL-PINN imposes an effective flexural-wave physics constraint consistent with the experimentally observed A0-dominated dispersive response. Experiments on aluminum plates demonstrate that the method reconstructs high-fidelity wavefields from only three signal averages (SNR ≈ 6 dB), achieving quality comparable to 300-average full scans with a substantially reduced pulse-count-limited acquisition burden. Compared with the supervised U-Net baseline under the tested random temporal split, the KL-PINN achieved higher reconstruction fidelity from sparse low-SNR inputs without requiring dense high-SNR labels for every frame. Furthermore, the accumulated KL-PDE residual provides a label-free scattering-footprint indicator associated with defect-induced wave interaction. These results indicate that physics-informed learning can reduce the acquisition burden in the tested thin-plate, full-optical LOU configuration. The experimental datasets, together with the KL-PINN implementation, are made publicly available to support further research. https://github.com/wangbaoding0816/Laser_Ultrasound_Reconstruction_PINN.