Leveraging physics-informed neural networks for wavefield analysis in laser ultrasonic testing

Laser Ultrasonic (LU) technology has emerged as a pivotal non-destructive testing method, offering a unique capability to visualise ultrasonic wavefields and identify defects without causing structural damage. However, challenges arise in certain testing scenarios where direct laser irradiation of the sample surface is hindered, resulting in incomplete LU wavefield datasets. This limitation poses a significant obstacle in accurately assessing material integrity and defect detection. This paper explores the application of Physics-Informed Neural Networks (PINNs) for LU wavefield reconstruction and prediction. PINNs are employed to reconstruct wavefields from incomplete data and predict wavefield behaviour at different time instances. Results demonstrate PINNs' effectiveness in accurately reconstructing wavefields, with correlation coefficients exceeding 0.94 between reconstructed and actual wavefields. Additionally, PINNs show promise in predicting LU wavefield data, albeit with slightly reduced accuracy beyond the training range. Moreover, PINNs effectively reduce noise in wavefield data, enhancing clarity and reliability. This study lays groundwork for further exploration of PINNs in LU defect detection.