Nondestructive evaluation measurements of large or complex structures leveraging full-field response maps to a prescribed ultrasonic excitation often exhibit distortion and scaling effects as a result of the inspection surface’s local position relative to the scanning laser-based acquisition system. To address these limitations, this study presents a deep-learning approach for estimating local material thickness from ultrasonic measurements of structural elements positioned at arbitrary orientations. The proposed model combines the full-field quadrature components of the measurement with an additional input for range data, such as that which would be collected with a LiDAR, to provide the network spatial context. Additionally, the model incorporates deformable convolutional layers which project locally onto the surface of the inspection specimen to improve the network’s robustness to different geometric transformations. To train the network, a pipeline was created to simulate acoustic steady-state excitation spatial spectroscopy measurements and depth field measurements of aluminum alloy plates with random variations in thickness and orientation. The model was evaluated on both simulated and real-world measurements, exhibiting better performance than other processing methods over a variety of scanning positions.
Maxwell et al. (2025) studied this question.