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

Area of Science:

• Non-destructive testing
• Wave physics
• Machine learning

Background:

• Full-optical laser ultrasonics (LOU) offers non-contact wavefield measurements but suffers from low signal-to-noise ratios (SNR) and long scanning times.
• Conventional interpolation and data-driven networks struggle with physically consistent field recovery under these limitations.

Purpose of the Study:

• To investigate a Kirchhoff-Love-plate-constrained physics-informed neural network (KL-PINN) for sparse wavefield reconstruction in LOU.
• To assess KL-PINN's capability for defect indication under low-SNR and sparse-sampling conditions.
• To reduce the acquisition burden in full-optical LOU for thin-plate analysis.

Main Methods:

• Embedding a Kirchhoff-Love thin-plate equation into a neural network to impose physics constraints.
• Experimental investigation using aluminum plates with sparse, low-SNR LOU data.
• Comparison with a supervised U-Net baseline for reconstruction fidelity.

Main Results:

• KL-PINN reconstructed high-fidelity wavefields from only three signal averages (SNR ≈ 6 dB).
• Reconstruction quality matched 300-average full scans, substantially reducing acquisition burden.
• KL-PINN outperformed U-Net in fidelity from sparse, low-SNR inputs without dense labels.
• Accumulated KL-PDE residual served as a label-free indicator for defect-induced wave scattering.

Conclusions:

• Physics-informed learning, specifically KL-PINN, effectively reduces the acquisition burden in thin-plate LOU.
• The KL-PINN framework enables high-fidelity wavefield reconstruction and defect indication from limited data.
• Publicly released datasets and implementation support further research in physics-informed wavefield reconstruction.