Quantification of microcracks by physics-guided neural network-assisted nonlinear ultrasonic technique

Area of Science:

• Materials Science
• Non-Destructive Testing
• Artificial Intelligence

Background:

• Quantifying microcracks smaller than ultrasonic wavelengths is challenging for traditional methods.
• Linear ultrasonic techniques face limitations due to wavelength-to-defect size ratios.
• Nonlinear ultrasonic techniques (NUT) struggle to isolate multiple damage parameters.

Purpose of the Study:

• To develop a physics-guided convolutional neural network (PGCNN) for quantitative microcrack characterization.
• To enable effective decoupling of crack length and width using nonlinear Rayleigh wave signals.
• To enhance the accuracy and robustness of microcrack assessment in non-destructive testing.

Main Methods:

• A physics-guided convolutional neural network (PGCNN) was developed, integrating a phenomenological damage indicator (DI) model.
• The DI model, based on crack aspect ratio, was embedded into the multi-task learning loss function.
• Experimental validation was conducted on torsion shaft specimens with varying microcrack geometries.

Main Results:

• The PGCNN demonstrated superior robustness and generalization compared to conventional convolutional neural networks.
• Effective decoupling of crack length and width was achieved from single nonlinear Rayleigh wave signals.
• Significant improvements were observed under data-poor conditions (40% training data): >15% R² increase and >33% MAE reduction for crack dimensions.

Conclusions:

• The proposed PGCNN-assisted NUT offers a powerful approach for microcrack quantification.
• Physics-based regularization enhances the network's consistency with nonlinear ultrasonic mechanisms.
• This method shows promise for reliable microcrack characterization in challenging scenarios, including data scarcity.