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

Area of Science:

• Materials Science
• Mechanical Engineering
• Artificial Intelligence

Background:

• Non-destructive evaluation (NDE) is crucial for metallic component integrity.
• Surface defects significantly impact fatigue life and performance.
• Laser ultrasonics offers non-contact inspection but faces challenges in wavefield reconstruction and defect identification from limited data.

Purpose of the Study:

• To develop a physics-informed neural network (PINN) for characterizing sub-millimeter surface defects in metallic components.
• To enable accurate forward and inverse modeling of laser ultrasonic wavefields.
• To reconstruct complete wavefields and estimate wave velocity for defect analysis.

Main Methods:

• Proposed a defect-parameterized physics-informed neural network (DP-PINN).
• Embedded defect parameters into elastodynamic equations for wavefield reconstruction.
• Simulated four defect cases and analyzed six practical scenarios with varying data sparsity and prior knowledge.

Main Results:

• Achieved defect characterization and full wavefield reconstruction using only 0.42 MB of measurement data.
• Demonstrated consistent defect detectability across diverse defect types.
• Obtained a mean Intersection over Union (IoU) of 0.387, indicating quantitative accuracy.

Conclusions:

• The DP-PINN effectively characterizes sub-millimeter surface defects in metallic components via laser ultrasonics.
• The method enables accurate wavefield reconstruction and defect localization with limited data.
• This approach enhances non-destructive evaluation capabilities for ensuring structural integrity.